# EtherGold (EG) version changelog Format: `MajorFeature.MinorFeature.MajorFix.MinorFix` (product BOM). **EG_VERSION** is the whole comm aggregate (HTTP, listen, pogcc-com, Python, OTA). Any of those component changes bumps this stamp — it is what OTA and `/health` show. Do **not** ship HTTP letter stamps as a second product. **UARTD_ABI_VERSION** (pog422_ipc.h `POG422_EG_VERSION`) is the last uartd/inject ELF rebuild. `update.sh` rebuilds those binaries when they ≠ ABI, not when Python/HTTP moves. XR driver may still fall in `[EG_COMPONENT_MIN_VERSION, EG_COMPONENT_MAX_VERSION]`. ## 2026-09-15 — EG **4.1.0.21** (`config-eg-busy-unlock`) `/config` Run EG / Stop EG no longer stick permanently disabled after a hung Update Now / promote / HTTP bounce. `field_config_ui.html` keeps a shared `egBusy` + `egRunLocked` sync (`syncEgRunButtons`), clears busy on login / send-to-login / OTA refresh failure, and applies `CHECK_TIMEOUT_MS` to EG Update Now. Collector lock (`run_eg_allowed === false`) still disables both buttons. UARTD ABI unchanged (**4.1.0.11**). ## 2026-09-15 — EG **4.1.0.20** (`fw-threshold-env-migrate`) `install_fixture_bridge_env.sh --ensure` (already called by `update.sh`) now idempotently rewrites leftover Phase 2 fireworks bench gates in `~/EtherGold/fixture-bridge.env`: `FIXTURE_FIREWORKS_WIN_THRESHOLD_USD=4[.00]` → `50.00`, and matching `FIXTURE_EARLY_WIN_{SLOT,KENO}_THRESHOLD_CENTS=400` → `5000`. Intentional overrides (e.g. `10.00`) are left alone. Missing keys stay missing — code default is already $50 from **4.1.0.19**. UARTD ABI unchanged (**4.1.0.11**). ## 2026-09-15 — EG **4.1.0.19** (`fireworks-threshold-50`) Phase 2 early-win fireworks default gate is now **$50.00** (was $4.00 bench default) for **all** Phase 2 cash games — slots (970C) and keno (TouchEasy 9309 / SuperBall 930A). Same `FIXTURE_FIREWORKS_WIN_THRESHOLD_USD` / enable-script `--usd` override still works for a lower bench gate. `run_bridge_jun11_good.sh` and the Phase 2 enable scripts now default to `50.00`. Phase 1 progressive (progwinval) unchanged. UARTD ABI unchanged (**4.1.0.11**). ## 2026-09-15 — EG **4.1.0.18** (`manual-listen-health`) A manual listen bench no longer reads as a broken field seat. `/health` uses `systemctl is-enabled ethergold-collector.service` as the oracle for who is *supposed* to own listen: unit **disabled** + `pog422_listen.py` up = `manual_listen`, `worker_owner: manual`, and neither `collector_inactive` nor `supervisor_down` is a fault (the listen supervisor belongs to the systemd seat). Unit **enabled** but inactive is still a hard fault, and `listen_down` still fails either way — a disabled unit with nothing running is not a healthy bench. `/health` adds `collector_enabled`, `manual_listen`, `worker_owner`; `services.collector` adds `enabled` / `worker_owner`. Health and dashboard Collector tiles show gray **manual** ("unit disabled — run_ethergold.sh owns listen") instead of red **down**, and the Listen PID tile reads `owner manual (run_ethergold.sh)`. UARTD ABI unchanged (**4.1.0.11**). ## 2026-09-15 — EG **4.1.0.17** (`run-eg-buttons`) `/config` **Start bridge** / **Stop bridge** renamed **Run EG** / **Stop EG** — they call `run_ethergold.sh`, which branches on `ETHERGOLD_ROLE`, so a listen seat starts listen (`run_listen_prod.sh`, or `run_listen_jun11_good.sh` when `FIXTURE_PROD=0`), not a bridge. Both buttons now lock out when `ethergold-collector.service` owns the worker: a second listen would contend for uartd / Pi HS BCM9, and `pkill pog422_listen.py` returns in `RestartSec`. `deploy_status()` adds `role`, `collector_active`, `worker_owner`, `run_eg_allowed`; the card hints Recover stack instead. Bench seats with no systemd owner (e.g. pos5 `.148`, role listen, `FIXTURE_PROD=0`) keep both buttons live. UARTD ABI unchanged (**4.1.0.11**). ## 2026-09-15 — EG **4.1.0.16** (`ota-channel-ui`) Show OTA channel on field UIs: `/EtherGold/config` Software OTA, `/EtherGold/health`, and `/EtherGold/status` dashboard. Red **OTA channel: /dev** when `OTA_CHANNEL=dev` (SGVI/dev); green **OTA channel: /prod** for field (`OTA_CHANNEL=field`). `/health` JSON and `/EtherGold/api/runtime` now include `ota_channel`. UARTD ABI unchanged (**4.1.0.11**). ## 2026-09-15 — EG **4.1.0.15** (`wire-bank-byte2`) Wire **Bank Number** is byte 2 of every SOH frame (`SOH|opcode|BANK|TID…`). `PacketMeta.bank_number`, `packet_complete` / `eth_rx` / `eth_relay` / `inject_echo` / SCR diag now stamp `bank` (+ `local_bank` / `bank_ok` when configured). Foreign-bank frames still land in captures but skip UART hooks, eth→UART inject, and `progressives.csv` (0x0A). Local bank from `EDGE_BANK` / `FIELD_BANK` / `BANK_NUMBER` / cabinet identity; unset (0) = accept all. UARTD ABI unchanged (**4.1.0.11**). ## 2026-09-15 — EG **4.1.0.14** (`uartd-own-not-contention`) Listen UART contention probe no longer treats **own** `pog422_uartd` as `foreign_pid_holds_port` when `POG422_UARTD=1` / `--uartd` (hat_listen starts uartd, then Python attaches via socket). Fixes Mini OHD collector restart loop and frozen `progressives.csv` (topbox33). Tip/witness still refuse a *foreign* collector unit when not `prod_mode`. UARTD ABI unchanged (**4.1.0.11**). ## 2026-09-14 — EG **4.1.0.13** (`usb-replug-job-clear`) Status USB cycle 2/2 modal loop: `finish` restarted HTTP mid-request so `apply-job.json` never left `await_replug` (Reconnected clicked many times; `usb_replug` itself already succeeded). Persist `finishing` before finish; skip HTTP restart under `WEB_APPLY`; clear job on success; Dismiss calls `/apply/usb-cancel`. UARTD ABI unchanged (**4.1.0.11**). ## 2026-09-14 — EG **4.1.0.12** (`hat-status-led`) POG-COM-PI-HAT **LED1** (BCM2, active-low): EtherGold POST + heartbeat. `hat_status_led.py` / `hat_status_led.sh` — N-count POST (100 ms pulses), healthy 250 ms, degraded 150 ms + soft-fault codes. `hat_bridge.sh` phases 1/2/4; bridge Python takes GPIO for 5–8 and healthy. Hardening: wait for `ttyXRUSB*` (phase 2), uartd start/connect retries (3/4/7). Env `EG_STATUS_LED_GPIO=2`, `HAT_STATUS_LED=0` disables. Requires `dtparam=i2c_arm=off`. UARTD ABI stays **4.1.0.11**. ## 2026-09-13 — EG **4.1.0.10** (`fixture-dev-rotated-inotify`) Fixture `/dev` only: inotify watcher offloads stamped capture under `~/EtherGold/dev/{data,logs}` to `DROPBOX_DEST`, then deletes local. Hard-refuses `/prod`. Scripts: `offload_fixture_dev_rotated.sh`, `watch_fixture_dev_rotated.sh`, `install_fixture_dev_rotated_offload.sh` (watch service + 30 min safety timer). Requires `inotify-tools`. UARTD ABI stays **4.1.0.9**. ## 2026-09-13 — EG **4.1.0.9** (`w2-hs-already-active-fall`) Pos4 overnight MRT (970D/`37BC` @ 05:44:43Z): deferred `81` on copper during local `89` Window 2; empty harvest; same bytes scooped later on next `0A` echo. Window 2 no longer waits up to 5 ms for HS IN *idle* after Window 1. After W1, HS should already be inactive; W2 starts µs later. If HS is **already active** at W2 entry, treat that as the in-flight `81` — drain until fall and collect FIFO (do not hunt a new rise). Idle at entry → wait rise → drain → fall flush. `UARTD_ABI_VERSION` / `POG422_EG_VERSION` → **4.1.0.9** (uartd rebuild). ## 2026-09-12 — EG **4.1.0.8** (`ohd-irq-affinity-quiet`) `update.sh` printed red `Job for ohd-irq-affinity.service failed` during the CPU-affinity step. Process pins were fine; the oneshot failed because systemd ran `apply_irq_affinity.sh` from the Dropbox tip (CRLF `sed -i` / `set -e`). Fix: install LF copies under `/usr/local/lib/ohd/`, CRLF heal via mktemp (no in-tree rewrite), `set +e` + always `exit 0`, and quiet `systemctl start`. UARTD ABI stays **4.1.0.7**. ## 2026-09-12 — EG **4.1.0.7** (`pi-hs-gpiochip-clock-de`) SCR `de_hold` / `pre_bulk` fluff on Pi HS: 1. **Hot path GPIO:** `pog422_pi_hs.c` uses gpiochip v2 ioctl (same as `pog422_gpio_hs.py` / SuperBall). `pinctrl` `system`/`popen` is init/fallback only — each fork was ~1–3 ms per DE set / IN sample. 2. **`de_hold`:** Pi HS waits `air+margin` with one `nanosleep` — no 100 µs HS-sample loop. Peer HS never asserts during SCR TX (Master answers after we drop DE); the loop only stretched hold past deadline via pinctrl tax. `UARTD_ABI_VERSION` / `POG422_EG_VERSION` → **4.1.0.7** (uartd rebuild). ## 2026-09-12 — EG **4.1.0.6** (`install-pi-expected-from-version`) Hotfix: `install_pi.sh` still hard-coded `EXPECTED_VER=2.2.2.11` after XR **2.2.2.12** built cleanly → `FATAL: built … version=2.2.2.12 want 2.2.2.11`. Now reads `EXPECTED_VER` from `xr-usb-serial-pog422/VERSION`. UARTD ABI stays **4.1.0.5**. ## 2026-09-12 — EG **4.1.0.5** (`ext-de-no-wire-sleep`) SCR inject wall was ~26 ms with idle_wait=0 because Pi HS paid air **twice**: 1. **XR 2.2.2.12:** `INJECT_TX` `EXTERNAL_DE` now returns after bulk like `MANUAL_DE` — no kernel `usleep(air+margin)`. Prior tip slept ~air inside the ioctl (`bulk_wall≈air`, `tx_us≈100 µs`) then C waited again. 2. **inject C:** Pi HS DE release is one air+margin wait (BCM sample / sleep); no `WAIT_TX_EMPTY` USB while DE high. 3. **uartd:** drop post-inject `gpio7_rx` on Pi HS (XR GPIO7 N/C; that USB control was ~6 ms of SCR `uartd_post_burst`). `UARTD_ABI_VERSION` → **4.1.0.5**. XR floor → **2.2.2.12**. ## 2026-09-12 — EG **4.1.0.4** (`scr-allow-slave-retry`) Spoke UART→ETH must **not** suppress Slave copper ``03`` retries while awaiting Master ``0B``. 4.1.0.2/3 ``scr_uart_eth_inflight`` blocked star re-ask for 220 ms after the first eth_tx — wrong when the glass has not seen ``0B`` yet (pos2 soak regression). Now only skip eth re-TX when same-TID ``0B`` is **queued or just injected**. Hub eth→UART same-TID coalesce unchanged. ## 2026-09-12 — EG **4.1.0.3** (`scr-coalesce-hook-sig`) Hotfix: `_on_packet_complete_scr_coalesce` used the wrong hook signature `(pkt, meta, …)` vs required `(meta, pkt, packets_in_window, …)`. Every copper `packet_complete` raised and **uart_to_eth went silent** after Dropbox picked up 4.1.0.2 without a formal stop/sync. Also skip `0B:TID` keys in `_scr_awaiting_reply`. ## 2026-09-12 — EG **4.1.0.2** (`scr-latency-shave`) SCR latency shaves after 4.1.0.1 unicast soak (retries remained ~20–40%/round). - **Same-TID SCR coalesce** (`SCR_COALESCE_MS=220`): drop eth→UART duplicate `03` while pending / in-flight / recently answered `0B`; same for duplicate `0B`; spoke UART→ETH suppresses re-TX of `03` while awaiting reply. - **`INJECT_PRIORITY_SCR=12`** beats soft `0A` / other STOM (10). - Soft progressive **`0A` defers** while any SCR awaits Master `0B` (UART budget). - `UARTD_ABI_VERSION` stays **4.1.0.0**. ## 2026-09-12 — EG **4.1.0.1** (`scr-tid-unicast`) SCR (`0x03` / `0x0B`) is **not** a bank broadcast. Remove ETH multi-spoke flood. - Mapped SCR: TID→gretap **unicast** (same as cash/skill). - Hub unmapped SCR: `[]` (nowhere) — never flood all spokes. - Spoke unmapped SCR: uplink flood still (`None`) so Local Net Test reaches Master. - Inject priority still treats SCR as STOM-class (queue order unchanged). - `UARTD_ABI_VERSION` stays **4.1.0.0** (Python/routing only). Soak driver: foreign-TID `0B` injects on spokes should drop to ~0 when TID map is warm. ## 2026-09-12 — EG **4.1.0.0** (`pi-hs-window-harvest`) Major refactor: Pi HS inject harvest is two HS windows, not timer peels. - **Window 1:** assert HS OUT + TX `89` → drain while our HS active (self-echo only) → **release HS OUT** so T340+ Slave can take the bus. - **Window 2:** wait HS IN rise → drain UART **only while HS IN active** → on fall flush once and parse local `81`. No post-inactive imm/post_de hunt. - UART-HS path keeps legacy imm/post_de peels. - Bridge: orphan RX + empty-harvest reinject **default off** (`MRT_X1_ORPHAN_DIAG`, `MRT_EMPTY_HARVEST_REINJECT`); progressive `0A` remains deferred until `0D` under `local_game_txn_exclusive` (WiFi star FIFO clean). - `UARTD_ABI_VERSION` / `POG422_EG_VERSION` → **4.1.0.0** (uartd rebuild required). ## 2026-09-12 — EG **4.0.0.73** (`uart-port-contention`) Tip/workshop listen must not share XR UART with prod collector. Dual-open can succeed while collector drains RX — HS LEDs look alive, tip sees `window_bytes=0` (looks like dead UART). - `probe_uart_port_contention` / `session_start` crumbs (holders + units) - `pog422_listen` **FATAL** when collector/bridge active or foreign PID holds `ttyXRUSB*` (override `ALLOW_UART_SHARE=1`) - `hat_listen.sh` early refuse if `ethergold-collector` active - `uartd_common` diagnose names collector ## 2026-09-12 — EG **4.0.0.72** (`boot-config-guard`) **Desktop/OS outranks HS-idle.** `.239` lost HDMI after hourly `install_hs_out_idle.sh` truncated `/boot/firmware/config.txt` to only the DYZINETEK_HS_OUT_IDLE gpio block (253 B) — Bookworm KMS (`dtoverlay=vc4-kms-v3d`) gone, no `/dev/dri`, lightdm crash-loop. Fix: - `EtherGold/sync/boot_config_guard.py` — health gate (KMS + `arm_64bit=1` + min size), golden backup on ext4 `/var/lib/ethergold/boot-config-golden.txt` (not boot FAT), atomic write + readback verify, refuse shrink vs golden, `repair` restores golden over a wiped config. - `install_hs_out_idle.sh` calls the guard for apply/remove; installs `ethergold-boot-config-guard.service` (sysinit, before display-manager) with authority to repair even if HS-idle is later uninstalled. - DF `sync/` mirrors guard + installer (delegates to EG when present). Never `sudo reboot`. Soft reboot on SEEED CM5 still needs hard power cycle. ## 2026-09-11 — EG **4.0.0.71** (`mrt-x1-orphan-rx-diag`) Diagnostic only (no product close/reinject change): after empty peel (`peel_src=0` / empty post-DE) arm a 300 ms spoke orphan RX watch (`post_x9_orphan_rx_arm` / `post_x9_orphan_rx` / `post_x9_orphan_x1_hit`) so a late Slave `81` can be proven in events. Hub logs `mrt_hub_master_x9` (incl. ~125 ms `is_retry`) and `mrt_hub_spoke_x1` (ETH `81` gap since Master `89`) for the same soak join. Disable with `MRT_X1_ORPHAN_DIAG=0`. ## 2026-09-11 — EG **4.0.0.70** (`empty-harvest-reinject-pending-expire`) Empty `89` peel (harvest=5) never reinjected: `pending_expire` AUTO_CLOSE_DONE IPC does not carry inject_echo, so `echo_ok` was always False and EG60 reinject never armed — while Master retried ~125 ms on hub copper and spoke deferred those retries behind `local_game_txn_exclusive`. Fix: keep echo/post_de bytes on `_mrt_risk_x9_diag` at inject; on miss recover `echo_ok` from diag so `_hand_open_after_auto_close_miss` lifts deferred / queues `empty_harvest_x9_reinject`. ## 2026-09-11 — EG **4.0.0.69** (`stale-handshake-adhoc-enroll`) Hub was reaping spokes as `stale_handshake` in a loop (CGS on ad-hoc join). Root cause: spoke re-enrolled every ~12s while WG `latest handshake` was still 0 (or only tens of seconds old — normal for WireGuard), so crypto handshake never completed; hub then dropped `handshake=0` peers after process-start grace. Fix: per-peer enroll grace on hub reap; spoke waits 90s for first handshake; re-kick threshold aligned with `EDGE_SPOKE_STALE_S` (~150s), not 12s. ## 2026-09-11 — EG **4.0.0.68** (`status-health-bridge-cards`) Status page Health strip was still listen-hardcoded (Collector red Down, Inject always painted as “0”). Now matches config: WiFi Hub/Spoke, Collector gray n/a on bridge, Inject path shows live software TX flag. ## 2026-09-11 — EG **4.0.0.67** (`health-cards-hub-spoke-inject`) Health polish for wifi-star: Collector is gray **n/a** (not red Down) on bridge profiles; Services show **WiFi Hub** / **WiFi Spoke** (or Side A/B) from node role. **Inject path** = software TX flag (`POG422_UARTD_INJECT` / `--uartd-inject`), detected from the live process — not a live packet counter. Listen still wants path off (0). ## 2026-09-11 — EG **4.0.0.66** (`health-mode-by-config`) Health cards follow field profile / `ETHERGOLD_ROLE` / `EDGE_MODE`: **listen** → inject=0, listen+collector+supervisor; **bridge** → inject=1, `run_ethergold` / `pog422_bridge`, no listen/collector faults. UI section title switches Listen ↔ Bridge. ## 2026-09-11 — EG **4.0.0.65** (`promote-tip-apply-defer-http`) Bugfix: Web Promote restarted `ethergold-http` inside `promote_staging.sh` **before** tip apply could rebuild inject/uartd, so Status stayed ABI MISMATCH after Promote. With `WEB_APPLY=1`, promote defers HTTP restart; tip-apply `finish` (or field skip path) owns the restart after rebuild/USB. ## 2026-09-11 — EG **4.0.0.64** (`web-promote-apply`) Web Status Promote/Update Now on **`OTA_CHANNEL=dev`** overlays Staging `prod/` onto **`~/EtherGold/dev` without `--delete`**, then tip-applies (rebuild inject/uartd when ABI/source requires it). USB driver cycles pause with browser **Disconnected / Reconnected** (mirrors `update.sh` Enter prompts) — never silent `SKIP_DRIVER=1`. Field channel stays prod-only promote. Status also shows tip/ABI/`tip_abi_ok`/`apply_phase`, plus **Start bridge**, **Stop bridge** (`SKIP_SYNC=1`), **Sync capture**. ## 2026-09-11 — EG **4.0.0.63** (`listen-until-x1`) Empty sync peel is **not** a definitive miss. Overnight EG60 SpinBall soak (~15% `auto_close_miss` on good `89` echo) reinjected while late `81` was still in flight — MRT / reinject storm class. 1. **uartd:** on good TX/echo but no x1 yet, do **not** send early `AUTO_CLOSE_DONE` NO_X1. Arm pending immediately. Imm peel ~100 ms; `AUTO_CLOSE_PENDING_NS` **120 ms** (Master re-x9 window). Pending **expire** sends `AUTO_CLOSE_DONE` miss once (`harvest=5` = pending_expire). 2. **Python:** sync `auto_close=None` → `status_0d_auto_close_pending` only (refresh await_x1). Miss / hand-open / EG60 reinject only on late `AUTO_CLOSE_DONE` (`via=pending_expire` when harvest=5). uartd ABI **4.0.0.10** — rebuild inject-shim on each Pi. **Soak checklist (vs overnight baseline `sessions/2026-09-11_1602-UTC_wlan-star-eg60-overnight-spinball-mrt2`):** - SpinBall AutoPlay on pos1–4 - Success: `auto_close_miss` ≪ prior ~15% of `89`; reinject storms rare - Hard `await_local_x1_timeout` down; glass MRT rate down vs baseline - Expect more `status_0d_auto_close_pending` + late ok / fewer premature miss ## 2026-09-10 — EG **4.0.0.62** (`ohd-proxy-df`) Public **`/ohd/*`** on ethergold-http `:31234` reverse-proxies to DyzineFrame localhost `:31235` (setup, modes, modes-status API). Status/home/config/health nav links. Clear 502 HTML/JSON when DF is down. Pair with DF **4.0.1.31** (always bind local HTTP when `ETHERGOLD_HTTP_EXTERNAL=1`). uartd ABI unchanged (**4.0.0.9**). ## 2026-09-10 — EG **4.0.0.61** (`fast-hub-reconnect`) Master-service power return: bank must talk again within **≤5 s** once the hub is up (spokes may stay running). 1. **WG PersistentKeepalive** default **5 s** (`EDGE_WG_KEEPALIVE_S`, was 25) — spoke conf, live `wg set`, hub peers, thumb/provision templates. Bridge start bumps live peers so old on-disk `= 25` does not wait a full 25 s. 2. **Spoke enroll**: no-hub poll **1 s**; retry backoff **0.5–5 s**; enroll UDP timeout **2 s**; **stay watching** after first success and emit `spoke_enroll_reconnect` when handshake is never/stale (>12 s). 3. **Bank WLAN announce** every **3 s** (`EDGE_BANK_ANNOUNCE_S`, was 15) so cold spokes learn hub uplink IP inside the budget. uartd ABI unchanged (**4.0.0.9**). ## 2026-09-10 — EG **4.0.0.60** (`empty-harvest-x9-retry`) Different angle after **4.0.0.59** porch still missed pos4 x1 (`post_de=0`, full ~129 ms peel, good `89` self-echo). Master same-seq `89` retries were deferred (`local_game_txn_exclusive` / `await_x1_peel`) until MRT — no second chance on copper. On `auto_close_miss` with **good self-echo** (delivery proved, harvest empty): 1. Arm `_empty_harvest_x9_retry_ok` so same-seq local x9 is **not** peel-deferred. 2. Lift any deferred Master x9 for that round onto the inject front, **or** 3. Queue one `empty_harvest_x9_reinject` of the missed x9. uartd ABI unchanged (**4.0.0.9**). ## 2026-09-10 — EG **4.0.0.59** (`pihs-x1-capture`) Tighten post-inject **local x1 capture** under Pi HS (residual MRT class after 4.0.0.58): good `89` self-echo + glass MRT (T340 sent `81`) but `post_de_len=0` / empty HS window. 1. **DE release** (`pog422_inject.c`): stop 1 ms `WAIT_TX_EMPTY` pad-loops that overshot air+margin (`de_hold` ~12 ms on misses while HS already live). When Pi HS + peer HS active past air time, free DE immediately. 2. **Peel porch** (`pog422_uartd.c`): imm first-byte gate == hard wall **60 ms** (was quiet-exit 20 / hard 45); post_de **45 ms**; HS x1 wait **60 ms**; auto_close pending **150 ms**; C FIFO grace **45 ms**. Re-assert Pi DE=0 at imm peel start. 3. Python local post-x9 FIFO grace **45 ms** (was 25). uartd ABI **4.0.0.9** — rebuild inject-shim on each Pi. ## 2026-09-10 — EG **4.0.0.58** Fix false mid-spin `mrt_risk_hand_txn_restart` (slots exposed hardest under Pi HS): Once local x1 is known for this play_sequence (`await_0d` / `x1_pending`), a duplicate same-seq local x0 must **not** clear await or bump `hand_gen`. Overnight soak showed early inject peel arming close, then a second `80`/`40` PC ~1 ms later orphaning the seal → `stale_hand_gen` / glass MRT. T340 will not re-wager until synth `0D`; that second PC is not a new game. Same-seq re-Play while still in `await_x1` (true miss recovery) still restarts. uartd ABI unchanged (**4.0.0.8**). ## 2026-09-09 — EG **4.0.0.57** Wlan-star soak follow-ups after 4.0.0.56 AutoPlay/MRT: 1. **ETH_PORTS gretap-only** for hub/spoke — never put `wlan*` on the AF_PACKET game plane (`run_bridge_edge_bench.sh` / `run_bridge_edge_prod.sh` / `pog_edge_env.sh`). `atlas_wifi_star` renders `EDGE_GAME_IF=none`. Link watch may still use uplink Wi‑Fi separately. Stops `eth_rx ??` / `bad_frame` junk from POGCOM RF being treated as game frames. 2. **MRT Pi HS `89`/`81` porch** — after early `peer_x1` harvest already queued `await_0d`, do not re-call `_mrt_risk_arm_after_local_x9` (false `mrt_risk_arm_skip` `await_0d_active`). On good closing 0D, clear `_local_hand_x9_gen` so the next local x0 is a cold hand (no false `mrt_risk_hand_txn_restart`). uartd ABI unchanged (**4.0.0.8**). ## 2026-09-09 — EG **4.0.0.56** Pos3 unhappy after 4.0.0.55 Pi HS tip: 1. **BCM8/9 were never configured as inputs** (`pinctrl get` showed `no | --`) so uartd Pi HS sense was blind — `packets.raw` empty, no `hs_active` / `packet_complete`. Fix: `pog422_pi_hs_ensure_pins()` sets IN `ip pu` and OUT `op dl` for machine+OIS. 2. **eth→UART thread crash** `select` on fd `-1` when hub `reopen_port` races enroll restore — filter invalid socks + catch ValueError. uartd ABI **4.0.0.8**. ## 2026-09-09 — EG **4.0.0.55** Bridge/uartd **Pi HS inject** for `POG422_HS_PATH=pi` (keep J3/J4 on PI — do not fall back to UART GPIO7 / 3 ms USB HS). When enabled: - **machine:** DE **BCM11**, IN **BCM9** (active-low) - **OIS:** DE **BCM10**, IN **BCM8** (active-low) `pog422_inject_burst` uses `EXTERNAL_DE` (no GPIO7) + pinctrl DE; `slv_in_live` / capture gating sense Pi IN. uartd ABI **4.0.0.7**. Rebuild `inject-shim` on each Pi (`make` in `inject-shim/`). ## 2026-09-09 — EG **4.0.0.54** **Spoke enroll writes wg0.conf via sudo + live hub Peer.** After hub accept, spokes failed `Permission denied: /etc/wireguard/wg0.conf` (root:700 dir) so no Endpoint / handshake while hub already had peers. `_write_wg_conf` uses `sudo install`; then `wg set` peer+endpoint live. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-09 — EG **4.0.0.53** **Hub `wg set` / gretap via ``sudo -n`` when bridge is topbox.** Enroll reached hub but `apply_failed` — Operation not permitted (bridge drops root after tunnel_up). Pos3 (`.187`) was still on **4.0.0.48** with root:700 spoke keys → Permission denied on `wg_private.key` (same class as earlier pos2/4). uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-09 — EG **4.0.0.52** **WG enroll is ad hoc anytime (no timed window).** Hub accepts new spokes whenever fleet HMAC + bank match; gates are crypto/bank, not `EDGE_OPEN_ENROLL` countdown. Optional lockdown: `EDGE_ENROLL_LOCK=1` restores allowlist/window behavior. Also export `EDGE_OPEN_*` / hub IP keys from `run_ethergold.sh` (were sourced but not exported). uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-09 — EG **4.0.0.51** **Spoke enroll uses `EDGE_BANK` (not empty GAMESTATS).** Bench seats with blank DyzineFrame `config.ini` had `bank_number=0`; enroll sent bank 0 while hub had `EDGE_BANK=1` → `enroll_denied` / `bank_mismatch` / no WG peers. Prefer Apply `EDGE_BANK` then identity. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-09 — EG **4.0.0.50** **Spoke enroll no longer dies on root:700 `/etc/ethergold/spoke`.** Bookworm `Path.is_file()` re-raises EACCES — wrap and fall through to writable keys. Apply `wlan_star_apply` chowns spoke key dir to `topbox` so HTTP-as-root Apply does not strand enroll. Ops residual: Jul-25 root keys on Duluth spokes were chowned live. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-09 — EG **4.0.0.49** **Hub bank-wlan discovery no longer waits for meters.** Hub publishes `discovery` at start and periodic `keepalive` every ~15s even with no `0x0A`, so late spokes learn `hub_uplink_ip` / enroll without typing Hub uplink IP. Prior: `maybe_keepalive` returned until the first meter — spokes stuck on `spoke_enroll_wait` / `no_hub_ip`. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-09 — EG **4.0.0.48** **/config Apply owns wlan-star WG (Pi SoT).** Writes `EDGE_SPOKE_ID` from Position Number (pos1 spoke→101), rewrites local `wg0.conf` on Apply, enroll returns `hub_wg_pub` and installs spoke Peer. Master=Hub is Node Role (any seat). PC `provision_wlan_star.ps1` demoted. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-09 — EG **4.0.0.47** **/config Live tracks the running HTTP tree.** `live_version` is the WorkingDirectory stamp (`/dev` or `/prod`), not only OTA `prod/`. Shows `tree /dev|/prod` and `outside Promote (OTA …)` when tip/SCP is ahead of the channel package. Staging/Rollback math still uses OTA package. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-09 — EG **4.0.0.46** **/config Node role was not actually read-only.** CSS `label:has(+ :read-only)` used the `:read-only` pseudo-class, which also matches `div` / sibling `label` / radios — so “Node role” and “Master Hub” got a false READ-ONLY badge. Scope to `input[readonly]` / `textarea[readonly]` (P3 GAMESTATS fields only). uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-09 — EG **4.0.0.45** **/config keeps Node Role selected after reload.** `fillForm` rebuilt role radios via `syncUi()` and cleared the checked hub/spoke radio. Preserve via `_pendingNodeRole`. Also recover `node_role` from applied `edge.env` (`EDGE_TOPOLOGY` / `EDGE_ROLE`) when JSON is blank. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-09 — EG **4.0.0.44** **HTTP tip alignment + banner tracks VERSION.txt.** 1. `floor_games_cache.HTTP_BUILD` was a hardcoded leftover (`4.0.0.39`) so `/health` kept showing .39 after tip **4.0.0.43**. Now reads tip `VERSION.txt` (no separate HTTP product stamp). 2. `update.sh` reinstalls `ethergold-http` from tip when the unit `WorkingDirectory` is still `~/EtherGold/prod` (stale `field_config_ui.html` — readonly TID). Fixture pos4 hit this. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-09 — EG **4.0.0.43** **POG422 TID is editable EtherGold identity — not P3-locked.** Hub/spoke/ listen type 4-hex TID on `/config`. P3 GAMESTATS panel is optional (DF/OHD); auto-grab on login only for listen profiles. Hub stamps `FIELD_TID` when set. 19CSTS / non-Atlas cabinets do not need a P3 for EtherGold. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-09 — EG **4.0.0.42** **Always park BCM10/11 output-low.** Identity Apply and listen/bridge startup drive OIS_OUT / M-S_OUT `op dl` even when `POG422_HS_PATH=uart`. UART protocol still uses XR GPIO1 for HS sense/drive — Pi pins are not toggled for handshake. Prevents TXS pull-up / DE-stuck when jumpers are on PI but config still says UART. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-09 — EG **4.0.0.41** **Dev capture rotate + rclone offload.** Bridge main loops (`uart_to_eth` / dual) call `try_rotate_debug_logs` + `idle_flush` on the stats cadence when HS is idle (listen already did). New `sync/offload_rotated_capture.sh` rclone-copies only stamped siblings (`events.jsonl.*`, `packets.raw.*`, …) to `DROPBOX_DEST/data/` then deletes local; live basenames stay open. Optional user timer `install_offload_rotated_capture_schedule.sh` (~10 min). `stop_fixture_session.sh` pre-offloads before live sync. Default rotate still **25 MiB** (`FIXTURE_DEBUG_LOG_MAX_BYTES`). uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-05 — EG **4.0.0.40** **FIXTURE_POS is routing only, never a run gate.** Field Apply already omits it from `ethergold.env`; collector systemd no longer requires `--fixture-pos` on ExecStart (empty env previously crash-looped `argument --fixture-pos: expected one argument`). Listen/bridge launchers soft-default to 1 with WARN. `resolve_fixture_pos()` + optional `--fixture-pos` on common CLI. Lab synth seat checks unchanged. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-05 — EG **4.0.0.39** **Listen-only OHD Mini profile.** New `listen_ohd_mini` (1920×360 / 24" panel): stamps DF `display_mode=MINI`, `test_mode_no_serial=True`, `disable_ads=True`, empty `progressives_source`. Operator chooses active games via `floor_games_source`=`local_p3`|`http` + `floor_games_http_host`. Grab GAMESTATS also refreshes `~/EtherGold/floor-games.json`. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-03 — EG **4.0.0.38** **Listen role never stamps inject=1.** `set_ethergold_role.sh` defaulted `POG422_UARTD_INJECT=1` before the role case, so `ETHERGOLD_ROLE=listen` commission wrote inject=1 into `/etc/ethergold/ethergold.env` (collector setdefault could not override an existing key). Listen now defaults and hard-forces inject=0. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. ## 2026-09-02 — EG **4.0.0.37** **Status CREDITS follows live Heltec telemetry.** `/EtherGold/status` read `machine-status.json` `pogcc`, which listen only refreshes on daisy packets. Listen down left pennies frozen (e.g. $0.55) after GAMESTATS had already zeroed Heltec. Runtime overlays `pogcc-telemetry.json`. Telemetry write force-publishes machine-status. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-09-02 — EG **4.0.0.36** **pogcc-com log is a capture file.** Heltec `/heartbeat` `/credits` and P3 GAMESTATS used to print only to journalctl, so `sync_to_dropbox.sh` never showed why Heltec bailed at 4 s. Now `~/EtherGold/dev/logs/pogcc-com.log` (line-buffered, rotate 1 MiB × 2). Logs `P3 open begin` *before* pyserial open, elapsed ms, `cmd=heartbeat done`, rx-queue drops. CLEAR_CAPTURE always keeps `pogcc-com.log*`. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-09-01 — EG **4.0.0.35** **Collector stop must release BCM9.** `ETHERGOLD_UARTD_PERSIST=1` plus `KillMode=process` and `start_new_session` used to leave `pog422_listen` running after `systemctl stop ethergold-collector`. Persist is uartd only; supervisor SIGTERM now always `stop_listen()`. `start_listen_background.sh` prints the leftover-child pkill when collector is already inactive. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-09-01 — EG **4.0.0.34** **GPIO9 EBUSY is another listen, not SPI.** `hat_listen.sh` used to auto-restart with no sleep on `GPIO9 already claimed`, so a second `start_listen_background.sh` while systemd `ethergold-collector` held BCM9 spun hundreds of FATALS/sec. Now: abort on `already claimed`; 2 s sleep on other auto-restarts; `start_listen_background.sh` pgrep `pog422_listen.py` / `hat_listen.sh` (not only `pog422_fixture_listen`); `stop_fixture_session.sh` same plus warns if collector `Restart=always` will bring listen back. EBUSY text no longer says SPI0. Persist trap prints only when UARTD is actually on. New collector installs use systemd `$FIXTURE_POS` from EnvironmentFile (live unit still has the install-time `--fixture-pos` until reinstall). uartd ABI still **4.0.0.6**. XR **2.2.2.11**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-09-01 — EG **4.0.0.33** **Listen flags live in the env file, not the command line.** `start_listen_background.sh` now sources `source_ethergold_env.sh` the same way `hat_listen.sh` / `run_listen_jun11_good.sh` already did, so workshop listen is `cd ~/EtherGold/dev && bash start_listen_background.sh`. Keys: `~/EtherGold/fixture-bridge.env` and `/etc/ethergold/ethergold.env`. Hostname `topbox6` maps to `FIXTURE_POS=6`. Launchers still force `POG422_UARTD_INJECT=0`. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-09-01 — EG **4.0.0.32** **GPIO-fall short frame: hold until known wire length.** Native listen `sessions/2026-09-01_1742-UTC_pos6-native-t340-listen-400030/` missed 2/462 `0x49` because USB delivered 18+26 and GPIO fall finalized the head (`split_empty`). Copper HS stayed high the whole 44 B. On fall, if the buffer is `SOH` + a `COMMAND_WIRE_LEN` opcode and still short, do not parse: `handshake_active` goes false (no DE/glue hold) and leftover tty bytes append for **1 ms** or until the next GPIO rise. Completeness is wire length, not interior `0x0D`. Recovers the 18+26 class. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-09-01 — EG **4.0.0.31** **CPU affinity is operator-disable.** `~/DyzineFrame/run/cpu_affinity.disabled` (HDMI Supervisor **CPU pin → Disable**, or `set_cpu_affinity.sh disable`) turns the 4-lane pin off so DyzineFrame uses all cores again — 2-core exclusive starved MP4; 0–2 still hurts fireworks. `cpu_layout.sh` / `install_cpu_affinity.sh` honor that flag (including under `sudo`, path `/home/topbox/...`) so `update.sh` cannot silently re-pin. Disable live-`taskset`s DF + EG pids onto all CPUs and removes systemd `CPUAffinity` drop-ins. `OHD_CPU_FORCE=1` ignores the flag. uartd ABI still **4.0.0.6**. XR **2.2.2.11**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-09-01 — EG **4.0.0.30** **Listen freeze: do not hold `handshake_active` on capture I/O.** 4.0.0.29 soak glued 170–214 B leftover UART and blocked retries / x1 for hundreds of ms because `emit` rotated, `packets.raw` / `inject_echo.jsonl` flushed, and hs_witness printed on every packet while GPIO was not sampled. Then the 500 ms gate was already open, so the next `0x40` landed ~10 ms after x1 and stomped Master `0x0D` ACK. Now: rotate/flush only when HS idle (`try_rotate` + `idle_flush`; `inject_echo.jsonl` rotates with events/packets); packet hooks run before disk; leftover UART does not set `handshake_active`; soak drains the 9th-bit tee without jsonl-emitting listen `wide_rx`; **25 ms close porch** after x1 before the next wager. XR **2.2.2.11**. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-09-01 — EG **4.0.0.29** **Idle UART is USB-late copper, not “bytes while DE inactive.”** Master cannot clock bits after he releases DE. If BCM9 already fell and XR delivers `05,49` a millisecond later, those bytes belonged to the window that just closed. Handshake idle drain now parses them (`uart_after_hs_fall` → `packet_complete`) instead of `bytes_discarded_inactive`. Overnight seq **125** was this class: discarded Master ACK, then three extra `0x40` retries. Does **not** wait on USB at GPIO fall (that was the 4.0.0.28 2 s jam). XR **2.2.2.11**. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. `inject_echo.jsonl` rotate is still later. ## 2026-09-01 — EG **4.0.0.28** **Empty-window HS finalize and own-echo wait must not USB-block.** Overnight pos6 soak (EG 4.0.0.27) proved BCM9 sees our BCM11 DE: every `0x40` has a 2–3 ms empty HS window, and `_handshake_finalize_window` still called `uart_rx_snapshot` (GET_REG / GET_RX_STATS). Four of five timeouts were us jamming the daisy (~2 s DE or discarded Master `0x49`). Handshake now drains via `uart_drain_no_usb` (`os.read` nonblock) and skips USB snapshot on empty windows (`verdict=skipped_usb_snapshot`). SuperBall `pulse_de_write` waits echo with `wait_own_echo_from_tty` so DE falls by the ~80 ms cap even if XR `in_waiting` would stall. Overnight proof: seq 89 discarded 214 B Master ACK; seq 105/228/212 held DE 2.1–2.4 s. XR **2.2.2.11**. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. `inject_echo.jsonl` rotate is still later (after Chris deletes the 4.0.0.27 landing). ## 2026-08-31 — EG **4.0.0.27** **Wide RX 9th-bit log + same-seq 0x40 retry 125 ms ×3.** UART already receives 9-bit (preciseflags + CS5); soak `ser.read` is 8-bit stripped. Player now enables the kernel wide RX tee (`SET_WIDE_RX`) and logs per-byte MARK/SPACE (`echo_9th` / `ninth_flags` on `inject_echo`, plus `wide_rx` events for listen). Missing Master `0x49`: re-TX the same seq `0x40` at 125 ms, three times, then `x9_timeout` (2 s cap if HS stays busy). Retries do not increment plays or move the 500 ms gate. XR **2.2.2.11**. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-08-31 — EG **4.0.0.26** **Dev soak logs sent vs echo hex.** SuperBall player compared echo to sent in RAM (`echo_ok`) but archives only kept verdict / length / last byte, so EVAL could not forensically diff a miss. `pulse_de_write` now calls `write_inject_echo` with `full_hex=True`: `sent_hex` / `echo_hex` / `match` / `match_kind` on `events.jsonl` `inject_echo`, plus `inject_echo.jsonl`. Prod bridge emit stays prefix-only. XR **2.2.2.11**. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-08-31 — EG **4.0.0.25** **Pi-HS soak holds BCM11 until own echo ends `0x0D`.** Skip-FIFO soak (4.0.0.24 / XR 2.2.2.11) dropped DE when `INJECT_TX` returned (~8.1 ms, `drain_wait` 7.80). Echo was 22/23 cropped (`F9`/`FF`/`F1`); Master LNT was HS blip only (23 TX / 0 TID **1002** `0x49`). Native `0x40` HS is ~11 ms. Player now keeps EXTERNAL_DE / skip FIFO / gpio7=0, waits `wait_own_echo` **before** `de.fall()`, then drops. Echo timeout (~57 ms) is still the jam cap. Log `de_hold=own_echo_eop`. Do **not** restore 18 ms 49→41 think-time. Do **not** set `POG422_UARTD_INJECT=1`. XR stays **2.2.2.11**. uartd ABI still **4.0.0.6**. ## 2026-08-31 — EG **4.0.0.24** **`update.sh` re-execs after pull** so the process uses the script that just landed. 4.0.0.23 Python pulled, then the still-running 4.0.0.22 script `FATAL`d `uartd 4.0.0.6 != tree tip 4.0.0.23` (the words `tree tip` are only in the old script). Same-turn unblock is `SKIP_PULL=1 bash update.sh` on the copy already on disk. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-08-31 — EG **4.0.0.23** **`update.sh` uartd check uses `UARTD_ABI_VERSION`, not `EG_VERSION`.** 4.0.0.22 pull on `.140` rebuilt uartd (4.0.0.6 in the ELF, correct) then `FATAL: after make, uartd version 4.0.0.6 != tree tip 4.0.0.22` and never reached XR **2.2.2.11**. Changelog already said ABI-not-tip; the inject compare still used tree tip (old pos5 “must match Python” guard). Driver MODULE_VERSION compare from 4.0.0.22 stays. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-08-31 — EG **4.0.0.22** **`update.sh` installs XR when tree `MODULE_VERSION` ≠ loaded.** 4.0.0.21 pull on `.140` left driver **2.2.2.10**: `_driver_needs_install` treated in-range `[EG_COMPONENT_MIN, EG_VERSION]` as done, and 2.2.2.10 still falls in `[2.2.2.10, 4.0.0.21]`. Soak `EXTERNAL_DE` needs **2.2.2.11**. Now install when `xr_usb_serial_common.c` `MODULE_VERSION` ≠ sysfs. Python-only stamps still skip driver (MODULE_VERSION unchanged). `EG_COMPONENT_MIN` stays **2.2.2.10** for listen. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-08-31 — EG **4.0.0.21** **Pi-HS soak DE closer to native air (skip FIFO get_reg).** Native T340+ `0x40` / `0x41` HS is ~11 / ~5 ms. Soak was ~35 / ~30 ms because BCM11 rose before `INJECT_TX` and the ioctl ran two TX_FIFO `get_reg` USB controls with DE already high. XR **2.2.2.11** `POG422_INJECT_POST_EXTERNAL_DE` (`0xFFFFFFFD`) skips those controls and does **not** assert GPIO7 (`MANUAL_DE` / `0xFFFFFFFE` still refused). Soak refuses to TX on XR older than 2.2.2.11 — that magic is a drain margin on 2.2.2.10. Listen still accepts driver **2.2.2.10** (`EG_COMPONENT_MIN` unchanged). `0x41` still waits HS inactive after Master `0x49`, then fires (do not add the native 18 ms think-time — progressive-win test is x1 on the wire before Master resends x9). Local `0x0D` synth stays later (Eth / WiFi RTT). uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-08-31 — EG **4.0.0.20** **Pi HS listen: kernel GPIO edges + no USB snapshot in the hot path.** Chris: slave `0x41` must drop HS before Master `0x0D` — the four 4.0.0.19 `window_glued` 41+0D rows were the same listen miss as 49+41, just a shorter dip (~1 ms extra in `byte_span` vs continuous air). Polling BCM9 once per tick still hid it because `uart_rx_snapshot` (FIONREAD + XR GET_REG) ran **every** open window. Now `GpioHandshakeMonitor` requests rise/fall events (1024-deep kernel buffer, non-blocking drain) and `handshake_capture_tick` splits windows on those edges. USB snapshot stays at window open/close only. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-08-31 — EG **4.0.0.19** **Listen must not block 20 ms on UART without sampling HS.** Pi HS listen (`handshake_capture_tick`) used `select()` with `HS_UART_INTER_CHAR_POLL_S = 0.02` after sampling BCM9 once. During that wait Master drops HS, the T340+ takes the bus, and `0x41` bytes land in the same buffer — `window_glued` / one `t_hs_rise` on both packets. Duluth SuperBall looked correct (155/155 separate, ~21 ms) because the x9→x1 gap was just longer than 20 ms so `select` timed out and GPIO ran. Native T340+ pos6 listen 17:41 AST glued **73/101**. Window still ends on GPIO inactive, not UART idle: non-blocking `in_waiting` drain, re-sample HS before the read. Constant removed. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-08-31 — EG **4.0.0.18** **P3 screen capture on field Home.** New Home tile **P3 capture** (`/EtherGold/p3-capture`). HTTP writes `~/DyzineFrame/run/p3_grab.request`; DyzineFrame **4.0.1.10** runs TARGAFILE on the P3 port it already owns (same as S key) and writes PNG + TGA under `~/p3-grabs`. Browser fetches the PNG with the UI session and **Save PNG on this device** (Downloads). HTTP never opens P3. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-08-31 — EG **4.0.0.17** **Sample HS immediately before SuperBall slave DE rise.** Pos6 11:09 soak (**4.0.0.16**, DE on ioctl return) was **111 / 109 / 109** with two `x9_timeout`s (seq 68, 102). Echo **220/220** last `0x0D` — not a crop. Those two misses had HS idle **368–391 ms** before DE; the remaining race is successful rounds that snatch the bus **0.92 ms** after `hs_inactive`. Loop already gated `hs_busy`, then built the frame, then `de.rise()`. Now `pulse_de_write` samples BCM9 twice (200 µs apart) immediately before asserting DE and aborts (`hs_busy_pre_de`) if another machine is driving. Skip does not increment seq / open a round / ACK. GPIO7 idle. `~/EtherGold/dev` only. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-08-31 — EG **4.0.0.16** **Soak echo must end 0x0D after INJECT_TX DE drop.** Chris: a premature DE drop could crop the packet; validate the local echo’s last byte is `0x0D`. DE still falls when `INJECT_TX` returns (do not hold for echo). After DE is low, classify echo: `echo_ok` (full frame + last `0x0D`), `echo_cropped` (bytes but not a complete EOP frame), `echo_timeout` (empty). Crop/timeout emit `error`. `session_stop` counts `echo_ok` / `echo_cropped` / `echo_timeout`. GPIO7 idle. `~/EtherGold/dev` only. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-08-31 — EG **4.0.0.15** **SuperBall slave DE drops when INJECT_TX returns.** Pos6 wide soak (14:13Z) was **207 / 201 / 201** with five `x9_timeout`s. Those five `0x40`s echoed fine (`usb_actual=84`) but Master never sent that seq’s `0x49` (three waits had **no HS for 2 s**). Successful `0x40`→`0x49` floor was **52 ms** — BCM11 held ~33 ms for userspace echo after the kernel had already waited ~8 ms wire. Fast Master results start as soon as we release DE; holding through echo collides with that path and Master does not retry. Echo is still logged (`inject_echo`). GPIO7 stays idle. `~/EtherGold/dev` only. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-08-31 — EG **4.0.0.14** **SuperBall slave TX is kernel INJECT_TX wide (MARK first / SPACE rest).** T340+ throws 8-bit SPACE-on-every-byte frames away; pyserial `write()` cannot switch parity between bytes without a gap. Soak now calls `XR_USB_SERIAL_INJECT_TX` (payload only; kernel packs `[data][0x01]` then `[data][0x00]`). BCM11 still owns DE. `post_margin_us` is never `MANUAL_DE` (`0xFFFFFFFE`) — that would assert XR GPIO7. Preciseflags + GPIO MODE/DIR after `open_serial(hs_source=pi)`. Workshop `~/EtherGold/dev` only. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-08-31 — EG **4.0.0.13** **SuperBall slave DE hold is own UART echo through 0x0D.** Last soak (pos6, 08-30 ~22:22) pulsed BCM11 ~12 ms after `flush()` (`POST_TX_MARGIN_US=400`) and the Master saw HS tick only. Hold DE until `ttyXRUSB1` RX matches the frame we just sent, including the terminating `0x0D` (do not stop on an interior 0x0D). Timeout still drops DE so a missing echo cannot jam the daisy. Drive BCM11 via gpiochip ioctl when the line is free; `pinctrl` remains the fallback. Workshop `~/EtherGold/dev` only. uartd ABI still **4.0.0.6**. Does **not** set `POG422_UARTD_INJECT=1`. ## 2026-08-30 — EG **4.0.0.12** **SuperBall slave soak stays in `/dev`.** `Promote-ToProductionStaging` / `build_prod.sh` were copying `superball_slave_*.py` and `run_superball_slave_soak.sh` into `/prod` (4.0.0.11). Those are workshop-only (`~/EtherGold/dev`, `ARM_TX=1`). Strip them from the prod fileset. uartd ABI still **4.0.0.6**. ## 2026-08-30 — EG **4.0.0.11** **SuperBall $20 slave soak (Pi HS, bench `.140`).** New `superball_slave_wager.py` / `superball_slave_player.py` / ledger: TID **1002** only (never Master TID), SuperBall `930A` $20 (80 credits) picks 1–5, pyserial `ttyXRUSB1`, BCM9 HS IN, BCM11 DE pulse (`op dh`/`op dl`), 500 ms start-to-start `0x40`s, `transactions.jsonl` + `jackpots.jsonl` (`progwin>0` + next `0x0A` game-11 `reset_ok`). TX requires `--arm-tx` / `ARM_TX=1`. Capture seat **`FIXTURE_POS=6`** / hostname **`topbox6`** → Dropbox `EtherGold/logs/6` (`install_sync.sh` maps `topbox1..6`; `stamp_pos6_capture.sh`). Does **not** convert GA 1–5, does **not** set `POG422_UARTD_INJECT=1`. uartd ABI still **4.0.0.6**. ## 2026-08-30 — EG **4.0.0.10** **Listen crash on first gameplay packet.** `observe_listen_current_game` called `map_wire_game_to_index` without importing it from `current_game_ipc` → `NameError` after the first `0x40`/`0x49` window, killing HS-gated listen mid SuperBall AutoPlay. Import fixed. uartd ABI still **4.0.0.6**. ## 2026-08-30 — EG **4.0.0.9** **Listen x0 checks DF HDMI game.** After each gameplay request (x0 wager), listen reads `df-displayed-game.json`. If missing, pre-boot, logo (`0`), or a different 1..16 than copper, it force-rewrites `current-game.json` even when the file already has SuperBall. DF **4.0.1.9** applies that rewrite. Does not rewrite on every x0 when HDMI already matches. uartd ABI still **4.0.0.6**. ## 2026-08-30 — EG **4.0.0.8** **GET /health lists EtherGold services.** Aggregate `version` stays `EG_VERSION`. Payload adds `product=ethergold` and `services` (http, listen, collector, pogcc_com). Existing `listen` block unchanged. Health HTML shows a Services row. uartd ABI still **4.0.0.6**. ## 2026-08-30 — EG **4.0.0.7** **EtherGold is one product.** HTTP listen/pogcc-com work from this morning (was letter **30-h**: keep Heltec ACM open / DTR; listen TID not spoke-only) now ships as **4.0.0.7**. `/health` `version` is `EG_VERSION`. OTA Staging shows **4.0.0.7**. uartd/inject ABI stays **4.0.0.6** (no UART rebuild). Does **not** start pogcc-com unless `pogcc_com_owner=ethergold`. ## 2026-08-30 — HTTP `ethergold-http-2026-08-30-h` (folded into EG **4.0.0.7**) **Do not close Heltec ACM.** 30-g hang watchdog closed the CDC fd at 20s (`pump hung — close CDC fd`). That drops DTR; Heltec reports the port closed even though `303a:0002` / `ttyACM1` stay on the bus. Closing `/dev/pts/1` every loop also stalled the pump. No hang abort, no 15s close/reopen, no pts `os.close`. Hold ACM open for DTR. Does **not** bump `EG_VERSION` / uartd. ## 2026-08-30 — HTTP `ethergold-http-2026-08-30-g` (still EG **4.0.0.6**) **Force `O_NONBLOCK` with `fcntl`; ACM `os.read` on a side thread.** 30-f still blocked in `os.read` (wchan `do_select`) before any hello; hang watchdog closed the fd from another thread and cdc_acm did not wake. Heltec theme applied then `heartbeat() Blocking took 4000ms`. Main loop only `Queue.get_nowait()` so status/`HB,` keep running if ACM read stalls. `F_SETFL` after `os.open`; hang abort 20s; drop leftover `/dev/pts` every loop. Does **not** bump `EG_VERSION` / uartd. ## 2026-08-30 — HTTP `ethergold-http-2026-08-30-f` (still EG **4.0.0.6**) **Do not `select()` on Heltec ACM.** 30-e opened ACM then the hang watchdog fired in 3s (`pselect` on this gadget ignores the timeout). CDC I/O is non-blocking `os.read`/`os.write` + `BlockingIOError`; Python `time.sleep(0.01)` on EAGAIN. Loop always ticks so `HB,` can be answered after theme. Does **not** bump `EG_VERSION` / uartd. ## 2026-08-30 — HTTP `ethergold-http-2026-08-30-e` (still EG **4.0.0.6**) **Heltec CDC is raw `os.open` / `select`, not pyserial.** After 30-d answered `HB,` and pushed Identity, ACM `ts` froze again. Hang watchdog logged `pump hung — cancel CDC I/O` but the process stayed in `pselect6(..., writefds=/dev/pts/1, timeout=NULL)` — not ACM. pyserial `cancel_read` cannot unblock a pts write. Heltec then applied theme and bailed on later heartbeats. Open ACM with `O_NOCTTY|O_NONBLOCK`, drop inherited pts fds, unit `StandardError=journal` (was inherit). P3 GAMESTATS still pyserial. Does **not** bump `EG_VERSION` / uartd. ## 2026-08-30 — HTTP `ethergold-http-2026-08-30-d` (still EG **4.0.0.6**) **pogcc-com hang watchdog is off-thread.** After 30-c answered the first `HB,` and pushed Identity, ACM `select` wedged (Heltec busy applying theme; gadget IN NAKs). Status `ts` froze; same-thread 15s watchdog never ran; Heltec then `heartbeat() Blocking took 4000ms` on every later poll. A daemon thread `cancel_read`/`cancel_write` + close if the pump loop stalls >3.5s. Identity is deferred until after `HB,` and written only when asset/theme/volume changed. Poll with `select` 50ms, not blocking `readline` on an empty ACM. Does **not** bump `EG_VERSION` / uartd. ## 2026-08-30 — HTTP `ethergold-http-2026-08-30-c` (still EG **4.0.0.6**) **pogcc-com must answer `HB,` within 4s.** Two hangs: (1) 15s CDC watchdog called `sudo usbreset 303a:0002` (no `-n`) and blocked on a password prompt; same landmine as DF — self-powered Heltec, cable VBUS open. Watchdog is close/reopen only; JTAG `303a:1001` uses `sudo -n`. (2) Identity `/asset` `/theme` `/volume` on **hello** filled TinyUSB RX before Heltec’s force heartbeat readline, and pyserial `write_timeout` default None wedged the pump so `HB,` never went out. Hello no longer downlinks; heartbeat writes `HB,` then identity. Serial `write_timeout=0.5`. Heltec **2802** IDE console: hello then `heartbeat() Blocking took 4000ms`. Does **not** bump `EG_VERSION` / uartd. ## 2026-08-30 — HTTP `ethergold-http-2026-08-30-b` (still EG **4.0.0.6**) **pogcc-com USB-pushes Identity.** When `pogcc_com_owner=ethergold`, DyzineFrame does not hold Heltec CDC. Apply already stamped `config.ini` `pog_cc_theme` / `pog_cc_volume` / asset, but `pogcc_com_service.py` (build `2026-08-12-c`) never sent `/asset` `/theme` `/volume`. It does now on hello and each heartbeat, so attract follows `/config` even if DF is down. Heltec **2801+** applies only if the value changed (NVS). Does **not** bump `EG_VERSION` / uartd. ## 2026-08-30 — HTTP `ethergold-http-2026-08-30-a` (still EG **4.0.0.6**) **Listen TID is not spoke-only.** `/config` Validate used to reject any `tid` on `listen_ohd` (`tid is only used for atlas_wifi_star slave_spoke`). That blocked Apply when leftover `FIELD_TID` sat in the hidden spoke field, even though listen current-game filter is the GAMESTATS 4-hex (`POG_TID`, Lucky Street **BD07**). Listen now accepts optional 4-hex TID; placeholders/`---`/junk are ignored. UI posts GAMESTATS `term_number` as tid on listen (not the hidden spoke box). Hub / wifi_bank_bridge leftover is ignored. Spoke still requires 4 hex. Apply stamps `POG_TID` when listen tid is 4 hex. Does **not** bump `EG_VERSION` / uartd. Not field-published. ## 2026-08-29 — HTTP `ethergold-http-2026-08-29-c` (still EG **4.0.0.6**) **Pi HS listen (workshop).** With J3/J4 on PI, XR GPIO1 is disconnected from THVD HS. Listen keeps XR UART RX (`ttyXRUSB`) and senses handshake on BCM9 `M-S_IN` / BCM8 `OIS_IN` (`pog422_gpio_hs.py`). Forces pyserial (`POG422_UARTD=0`) because uartd still samples XR GPIO1. Idle OUTs stay driven low. Does **not** inject. Does **not** bump `EG_VERSION` / uartd. Not field. ## 2026-08-29 — HTTP `ethergold-http-2026-08-29-b` (still EG **4.0.0.6**) **Pi HS OUTs idle-low on every Pi at boot.** BCM10 `OIS_OUT` and BCM11 `M-S_OUT` are driven low from firmware `config.txt` (`gpio=10=op,dl` / `gpio=11=op,dl`) and systemd `ethergold-hs-out-idle.service` (sysinit pinctrl). Older POG-COM-PI-HAT boards have those pads unconnected (safe). Rev E+ must never leave the Pi pin high-Z (TXS pull-up holds DE on). Installer `sync/install_hs_out_idle.sh` also runs pinctrl immediately (no reboot). Hooked from Promote, listen-stock, and DF harden. Independent of the `/config` UART vs Pi HS toggle. Does **not** bump `EG_VERSION` / uartd. ## 2026-08-29 — HTTP `ethergold-http-2026-08-29-a` (still EG **4.0.0.6**) **Deployment profile High Speed path** is its own toggle, not a new profile. `uart` (default) = HAT XR HS (J3/J4 UART). `pi` = Pi GPIO J19 (J3/J4 PI). Apply writes `POG422_HS_PATH` and, on a Pi with `pinctrl`, idle-drives BCM10/11 **low** for Pi HS (never high-Z). Does **not** change `POG422_UARTD` / inject / `FIXTURE_POS`. GPIO listen backend is not field yet. Does **not** bump `EG_VERSION` / uartd. ## 2026-08-28 — HTTP `ethergold-http-2026-08-28-c` (still EG **4.0.0.6**) **Pi Identity owns Heltec theme and volume.** `/config` Identity card has attract theme (1–5 Slots/Numbers/Rainbow/Fire/Quiet) and DS1881 volume (0–62). Apply stamps `config.ini` `pog_cc_theme` / `pog_cc_volume`. DyzineFrame **4.0.1.4** USB-pushes those plus `/asset` down; leftover Heltec `/asset` `/ip` are ignored. Heltec Setup (`POG_CC_2026082801`) is Wi-Fi only. Does **not** bump `EG_VERSION` / uartd. Not published unless asked. ## 2026-08-28 — HTTP `ethergold-http-2026-08-28-b` (still EG **4.0.0.6**) **Onsite enroll during POG-CC + Pi provisioning.** Identity Apply always upserts this cabinet’s Enrollment.csv row (including `listen_ohd_server`). Config/Status Location dropdown has **new site** (type Location + Path). POG-CC USB `/asset` upserts the same row when `config.ini` already has a Path. Ads/content sync can use Identity Path before the asset is in the Dropbox SGVI catalog. Pi still does not publish Dropbox SGVI. Workshop catalog row **21044** Lucky Street was added in `SGVI/Enrollment.csv`. Does **not** bump `EG_VERSION` / uartd. ## 2026-08-28 — HTTP `ethergold-http-2026-08-28-a` (still EG **4.0.0.6**) **Recover stack** on Config/Status: listen/collector first, wait for `progressives.csv`, then DyzineFrame `supervisor.cmd` restart. Lucky Street 08-28: DF 4.0.1.3 unlinked pre-boot CSV; Atlas still saw 0x0A; HDMI banner stayed until DF restart. Never `sudo reboot`. Never kill HDMI supervisor. Listen supervisor no longer parks 300 s after five CSV-missing bounces. Does **not** bump `EG_VERSION` / uartd. ## 2026-08-26 — HTTP `ethergold-http-2026-08-26-f` (still EG **4.0.0.6**) **Staging older than Live is Rollback, not pending.** LAN/workshop live ahead of SGVI (green disc + 03:00 would have downgraded). Stage parks the older tree into Rollback and clears `staging-pending`. Promote skips that downgrade unless `CONFIRM_DOWNGRADE=1`. Config copy matches. Pairs with DF **4.0.1.3**. Does **not** bump `EG_VERSION` / uartd. ## 2026-08-26 — listen early-win (still EG **4.0.0.6**) **Listen-only triggers early fireworks.** Field TopBox is a copper witness, not a bridge, but it already sees complete x9 frames. Phase 1 (progwin, including SuperBall `930A` / `0x49`) and Phase 2 keno/slots now run on `pog422_listen.py` via `fireworks-request.json`. Inject stays off. Listen defaults the `FIXTURE_EARLY_WIN_*` flags **on** (bridge stays opt-in). Phase 2 cash still needs local TID (`ETHERGOLD_LOCAL_WIRE_TID` / `POG_TID` / `ETHERGOLD_CURRENT_GAME_TID`). Does **not** bump `EG_VERSION` / uartd. ## 2026-08-26 — HTTP `ethergold-http-2026-08-26-e` (still EG **4.0.0.6**) **Restart DyzineFrame from Config/Status.** `POST /EtherGold/api/dyzineframe/restart` writes `~/DyzineFrame/run/supervisor.cmd` `restart` and clears `supervisor.hold` (same file HDMI Supervisor uses). Does **not** kill `dyzineframe_supervisor.py`. Does **not** `sudo reboot`. Does **not** bump `EG_VERSION` / uartd. Pairs with DF **4.0.1.1**. ## 2026-08-26 — HTTP `ethergold-http-2026-08-26-d` (still EG **4.0.0.6**) **Health copy: SDS Enter Game writes `current-game.json`.** Not copper gameplay, not 0x0A meters. Missing until Heltec SDS 0xA2 `/sds` (1..16) this boot is expected on attract. Does **not** bump `EG_VERSION` / uartd. ## 2026-08-26 — HTTP `ethergold-http-2026-08-26-c` (still EG **4.0.0.6**) **Formatted health page.** `GET /health` stays machine-readable JSON (boot_id watch, curl). HTML twin is `GET /EtherGold/health` (no login; polls the same JSON). Home tile + Config/Dashboard nav. Status dashboard has a Health card after login. JSON includes `"ui": "/EtherGold/health"`. Does **not** bump `EG_VERSION` / uartd. Does **not** Promote/Publish-Sgvi. ## 2026-08-26 — HTTP `ethergold-http-2026-08-26-b` (still EG **4.0.0.6**) **Missing `progressives.csv` is a listen fault, and listen comes back.** Collector supervisor no longer treats leftover `hat_listen.sh` as “still launching” forever (that blocked respawn after listen Python died). After 45 s grace it bounces listen if the CSV is absent/empty (max 5, then 300 s give-up). `GET /health` `listen.reasons` includes `progressives_csv_missing` so `listen.ok` is false. Does **not** bump `EG_VERSION` / uartd. Does **not** Promote/Publish-Sgvi. ## 2026-08-26 — HTTP `ethergold-http-2026-08-26-a` (still EG **4.0.0.6**) **Lucky Street listen hardening (workshop; unpublished).** `GET /health` now includes `listen` (collector, listen/supervisor PIDs, `ttyXRUSB*`, uartd sock, CSV/current-game age, inject must be 0). HTTP `ok` stays true if HTTP is up. Collector unit + drop-in: `Type=notify` `WatchdogSec=60`; supervisor sd_notify READY/WATCHDOG (stdlib, no python-systemd). `ota_restart_ethergold_listen` restarts `ethergold-collector` (`KillMode=process`); does not SIGKILL the supervisor or mention `ethergold-listen.service`. Does **not** bump `EG_VERSION` / uartd. Does **not** Promote/Publish-Sgvi. ## 2026-08-25 — EG hourly + 03:00 timers on promote (still EG **4.0.0.6**) **Listen-stock only installed DyzineFrame timers**, so Lucky Street `10.1.1.39` kept EG **4.0.0.4** with no `ethergold-stage.timer` / `ethergold-promote.timer`. `promote_staging.sh`, `stage_from_dropbox.sh`, rollback, and `prepare_listen_stock.sh` now run `sync/install_pi_deploy_schedule.sh` (idempotent). Update Now / Check for Updates also sets `XDG_RUNTIME_DIR` so `systemctl --user` works from `runuser -u topbox`. Does **not** bump `EG_VERSION` / uartd. Does **not** `sudo reboot`. ## 2026-08-25 — collector User=topbox drop-in (still EG **4.0.0.6**) **Listen-stock leftovers ran `ethergold-collector` as root**, so `~/EtherGold/logs/supervisor.log` was root-owned and `log_offload.sh` (topbox) aborted. `install_uartd_persist_dropin.sh` now also writes `User=topbox` (same drop-in as `KillMode=process`), chowns that logs dir, and restarts the collector **only when MainPID is not topbox**. Promote, rollback, and `log_offload.sh` all run it. Idempotent. Does **not** bump `EG_VERSION`. ## 2026-08-25 — HTTP `ethergold-http-2026-08-25-f` (still EG **4.0.0.6**) **Production log PUT uses curl + Basic auth.** Debian wput is FTP-only. `log_offload.sh` does `curl -T` with `~/EtherGold/sync/wput.env` (`WPUT_USER` / `WPUT_PASSWORD`, user **SGVIUpload**). Anonymous GET/OTA unchanged. Does **not** bump `EG_VERSION`. ## 2026-08-25 — HTTP `ethergold-http-2026-08-25-e` (still EG **4.0.0.6**) **Machine move updates Enrollment.csv.** Identity Apply upserts this cabinet’s Asset/Location/Path row in `Enrollment.csv` and `Enrollment.local.csv` (overlay still wins after an OTA pull of an older catalog). The Pi does not publish Dropbox SGVI. Does **not** bump `EG_VERSION`. ## 2026-08-25 — HTTP `ethergold-http-2026-08-25-d` (still EG **4.0.0.6**) **Identity Grab GAMESTATS.** `/config` P3 panel has a button that POSTs `/EtherGold/api/identity/p3-grab` with `force` (re-read even if TID is already filled). Still idle-window only — does **not** GAMESTATS while `current-game.json` shows recent play. Does **not** bump `EG_VERSION`. ## 2026-08-25 — HTTP `ethergold-http-2026-08-25-c` (still EG **4.0.0.6**) **Identity read-only vs editable.** `/config` groups TID, machine number, and Mode in a dashed “From P3 GAMESTATS” panel. CSS `:read-only` is the common treatment (dashed border, muted fill, READ-ONLY on the label). Operator fields keep a solid border and gold focus. Does **not** bump `EG_VERSION`. ## 2026-08-25 — HTTP `ethergold-http-2026-08-25-b` (still EG **4.0.0.6**) **Identity Location + log upload.** `/config` Identity card: Enrollment.csv Location dropdown and **Upload logs to pog-atlas** (default on). Apply stamps DF `config.ini` and, for a new asset at a known site, `Enrollment.local.csv`. **wput** to `:51709` `{Path}/Logs/` or `/orphan/{asset}/` if the asset is set but not enrolled. Unset asset skips IIS. rclone is not the log client. Does **not** bump `EG_VERSION` / uartd. **No pre-boot jackpot CSV.** Listen and `ethergold-http` unlink `progressives.csv`, `progressives-sync.json`, and `current-game.json` when those files were written before this OS boot. Lucky Street power-up had shown SuperBall at office ~$5k. Do **not** bind or redirect port 80 — HDMI shows the IP; `:31234` stays unpublished. ## 4.0.0.6 — 2026-08-24 — sgvi-ota **Promote restarts ethergold-http.** After Staging→live rsync, `promote_staging.sh` runs `ota_restart_ethergold_http` (`sudo -n systemctl restart ethergold-http`). Does **not** restart `ethergold-collector`. `SKIP_ETHERGOLD_HTTP=1` skips. Rollback does the same. Prevents new field HTML from talking to a stale HTTP process (login session vs Basic). First stamp with this OTA HTTP restart. ## 4.0.0.5 — 2026-08-24 — sgvi-ota **Login return URL.** Unauthenticated `/EtherGold/config` and `/EtherGold/status` redirect to `/?next=`. After Sign in, `/` navigates back to that allowlisted path (`/EtherGold/config`, `/config`, `/EtherGold/status`, `/status`). **HTTP session dies with the process.** Sign in mints an in-memory UI session (`HTTP_BUILD` `ethergold-http-2026-08-24-a`). Stored Basic is not enough after an HTTP restart or Pi reboot. Open config/status/home tabs watch `/health` `boot_id` and bounce to login when it changes. APIs reject stale session headers with `session_stale` (no browser Basic dialog). ## 4.0.0.4 — 2026-08-24 — sgvi-ota **Promote applies uartd persist drop-in.** After Staging→live rsync, `promote_staging.sh` runs `sync/install_uartd_persist_dropin.sh`: writes `/etc/systemd/system/ethergold-collector.service.d/uartd-persist.conf` (`KillMode=process`, persist env) and `daemon-reload`. Does **not** restart `ethergold-collector`. Needs `sudo -n`. `SKIP_UARTD_PERSIST_DROPIN=1` skips. `hat_listen.sh` / listen supervisor default persist; `uartd_common.sh` will not SIGKILL uartd unless `ETHERGOLD_UARTD_ALLOW_SIGKILL=1`. Rollback leaves the drop-in in place. First field stamp with this OTA systemd path. ## 4.0.0.3 — 2026-08-23 — sgvi-ota **Check for Updates.** `/EtherGold/config` (also `/config`) button rclone-pulls SGVI into Staging only — no promote, no listen restart. POST `/EtherGold/api/dyzineframe/check-updates` and `/EtherGold/api/ethergold/check-updates`. **HTTP_BUILD `ethergold-http-2026-08-23-f`:** Check for Updates greys OTA buttons then re-enables them in `finally` (DF Promote comes back as soon as the DF pull finishes; EG rclone no longer holds the DF row). Status refresh runs after buttons are live again. 210 s fetch abort so a hung rclone cannot leave the page stuck. **Field UI shell (`HTTP_BUILD` `ethergold-http-2026-08-23-e`):** `/` is login, then Home tiles for Config and Dashboard. JSON health moved to `/health`. Default login user SGVI (password in `field_config.py` / `field-ui.auth`, aligned to the field PSK already used by `leave_ap_join_wifi.sh`). `/config` loads `field-config.json` into the form. **Apply** (above Software OTA) writes `/etc/ethergold` env files with no reboot and no listen/collector restart. TopBox profile `listen_ohd` is DyzineFrame TBX 1920×1080 (machine hooks) **and** HTTP :31234 progressives (Win11 OHD can consume the same feed). New `listen_ohd_server` is that same HTTP path with no local display. ## 4.0.0.2 — 2026-08-23 — sgvi-ota **Field publish.** Same OTA machinery as 4.0.0.1. 4.0.0.1 stayed LAN-only on the St. Thomas Pis so field SGVI would not still be 16-c when timers come back. First `Publish-Sgvi.ps1 -Channel field` stamp for this train. ## 4.0.0.1 — 2026-08-23 — sgvi-ota **Major:** Field software OTA is rclone from Dropbox **SGVI** (not pog-atlas.com:51709). Chris publishes with `Publish-Sgvi.ps1 -Channel field` only. Pis hourly-stage `~/EtherGold-Staging`, promote at 03:00 America/St_Thomas (or `/config` on that Pi). `ROLLBACK.json` applies once by id; promote-hold blocks re-applying the broken stamp. Listen restart is numeric python PIDs — never `pkill -f`, never `sudo reboot`, UART `fuser` off unless `REBUILD_UART=1`. Includes **3.1.0.7** pog-updt-0a-pass. ## 2026-08-22 — listen-stock keeps live Wi-Fi; current-game IPC on **Stock:** `prepare_listen_stock.sh --keep-office-wifi` no longer deletes POGCOM (or the active wlan0 link). `--strip-wifi` still wipes all SSIDs. Env stamp includes `ETHERGOLD_CURRENT_GAME_IPC=write`. `SKIP_DF_SEED=1` skips Dropbox DF when LAN-seeding. ## 2026-08-21 — listen writes current-game.json **Listen:** `pog422_listen.py` publishes `~/EtherGold/current-game.json` from UART gameplay (same `CurrentGamePublisher` as the bridge). DyzineFrame’s watcher can then `sds_handler()` after Heltec flash / SDS-gap. Collector env defaults `ETHERGOLD_CURRENT_GAME_IPC=write`. Optional `ETHERGOLD_CURRENT_GAME_TID` filters one cabinet on a bank tap (otherwise last mapped wager on this HAT). ## 2026-08-21 — listen-stock boots collector **Stock:** `prepare_listen_stock.sh` enables `ethergold-collector` + `ethergold-http` on boot (`POG422_UARTD_INJECT=0`). It no longer leaves `/var/lib/ethergold/disabled` or waits for site apply. `install_collector_service.sh` merges env (keeps inject=0 / CHANNEL=both). ## 3.1.0.7 — 2026-08-16 — pog-updt-0a-pass **Fix:** Master **does** answer T340+ `0x02` (POG_UPDT). The reply is **`0x0A` @ 82** (`stom_update`) — same opcode as the meter poll. Hub flooded it; EDGE spokes **stashed it as `0a_latest_wins`** and never UART-injected it. Copper already has that frame on the wire. `0x0A` is now STOM-priority (never game-gate defer). Pos5 must run this rev (not only the hub). Includes **3.1.0.6** t340-cgs-config-pass. ## 3.1.0.6 — 2026-08-16 — t340-cgs-config-pass **Fix:** T340+ CGS / Transfer Master Config must reach Master UART (copper would). - Spoke **`0x02`** (and `0x01`/`0x04`) was ETH'd but hub **deferred inject** while any local close (`await_local_0d`) was open — and **dropped** under defer queue cap. Peer gameplay was allowed; config was not. - Hub **unknown-TID** UART→ETH of Master **`0x09`/`0x0E`** used `[]` (send nowhere). Unmapped STOM now floods; mapped still unicasts. - STOM is keep-score 95 if still deferred; stom inject is no longer gated off during local close (synth 0x0D stays P0). Includes **3.1.0.5** spoke-stale-reap. ## 3.1.0.5 — 2026-08-16 — spoke-stale-reap **Goal:** WLAN star hub drops dead / swapped spokes without a reboot. - Hub reaps WireGuard peers with no handshake for **`EDGE_SPOKE_STALE_S`** (default **180s**). - Removes `wg` peer + `eg-sNN` gretap, hot-drops the ETH port (`DirectionalEthRouter.remove_port`), updates `.active_spoke_gretaps`. - Replacement board with the same preferred spoke id steals the slot once the old handshake is stale. - Set `EDGE_SPOKE_STALE_S=0` to disable. Events: `hub_peer_dropped`, `hub_eth_port_removed`. Includes **3.1.0.4** pogcc-com-update-sync. ## 3.1.0.4 — 2026-08-12 — pogcc-com-update-sync **Goal:** Idempotent bank roll via `update.sh`. - `install_pogcc_com_service.sh` modes: `auto` (default), `--enable-now`, `--stop`, `--unit-only` - `auto`: refresh unit; enable only if DF `pogcc_com_owner=ethergold`; else stop+disable - `EtherGold/sync/update.sh` + `DyzineFrame/update.sh` call auto sync (`SKIP_POGCC_COM=1` to skip) - DF update ensures `pogcc_com_owner` key exists (default `dyzineframe`) - **2026-08-16:** `commission_prod_boot.sh` also installs **`ethergold-http`** (`:31234`) and `loginctl linger` so Production boot is automatic after commission (DF still starts from desktop autostart, not SSH). Includes **3.1.0.3** pogcc-com-df-local-http. ## 3.1.0.3 — 2026-08-12 — pogcc-com-df-local-http **Fix:** On office/collector Pis, `ethergold-http` owns `:31234` and DF skips Bottle — pogcc-com `/flag` + `/pogcc/hb-state` 404'd. Unit now targets `http://127.0.0.1:31235`; DF **2026-08-12-b** binds that localhost port when `pogcc_com_owner=ethergold`. Includes **3.1.0.2** pogcc-com-runtime-dir. ## 3.1.0.2 — 2026-08-12 — pogcc-com-runtime-dir **Fix:** `ethergold-pogcc-com.service` crash-looped (`PermissionError` on `/run/ethergold`). Add `RuntimeDirectory=ethergold`; status write falls back to `~/EtherGold/status/`. Includes **3.1.0.1** pogcc-com-service. ## 3.1.0.1 — 2026-08-12 — pogcc-com-service **Goal:** EtherGold owns Heltec USB CDC + P3 GAMESTATS for POG-CC COM (canary). MinorFeature bump: new `ethergold-pogcc-com` ownership path (DF becomes display-only when armed). - `pogcc_com_service.py` / `pogcc_com_ports.py` / `ethergold-pogcc-com.service` - Install: `bash install_pogcc_com_service.sh` (after DF `pogcc_com_owner=ethergold`) - Resolves Heltec by USB VID `303a` (prefers last ACM if dual); P3 via `/dev/pogcc` - Writes boot-arm + telemetry; `/sds` → current-game.json; `/flag` → DF HTTP - Status: `/run/ethergold/pogcc-com.status` - Pos5 canary: `POS5_POGCC_COM_CUTOVER.txt` Includes **3.0.2.14** df-except-eg-mask. ## 3.0.2.14 — 2026-08-10 — df-except-eg-mask **Goal:** Exclusive DF pin to core **2** (3.0.2.13) starved fireworks — sub‑1 FPS. **Fix:** DyzineFrame uses **`OHD_CPU_DF_SET=0-2`** (all cores except EG). EtherGold stays exclusive on core **3**. Live `update.sh` re-pins DF to the multi-core set. Includes **3.0.2.13** df-live-cpu-pin. ## 3.0.2.13 — 2026-08-10 — df-live-cpu-pin **Goal:** After tip install, live DyzineFrame was still free on cores 0–3 (old supervisor, not taskset-wrapped) while EG was pinned to core 3 — DF could still share EG’s core. **Fix (`install_cpu_affinity.sh`, driven by `sync/update.sh`):** 1. After drop-ins/IRQ oneshot, **re-pin live** `pog422_uartd`, bridge Python, `dyzineframe_supervisor.py`, and `DyzineFrame.py` to the correct cores. 2. If `~/DyzineFrame/cpu_affinity.sh` is missing, write a minimal helper so next runit/supervisor start uses taskset. Includes **3.0.2.12** txn-restart-local-x0. ## 3.0.2.12 — 2026-08-10 — txn-restart-local-x0 **Goal (pos5/witness51 MRT-clear thrash):** After hard MRT clear the glass restarts the full transaction (often same `play_sequence` 40). EG left half-finished await-x1 / late 41→0D from the abandoned attempt, so the restart was not clean and the cabinet re-latched MRT. **Fix:** 1. Local slave gameplay request (x0/40) **always** restarts the hand: abandons open close-wait **including same seq** (supersede was newer-only), force-closes same-seq trunk so `local_game_txn_open` can fire again, and runs existing same-seq close-done reopen (recovery immediate / ok settle). 2. **Hand generation:** each restart and missing-x1 abandon bumps a gen and unbinds prior x9; local x1 / uartd auto-close only seal when x9 gen matches current (stale late 41/0D after restart are dropped). Includes **3.0.2.11** cpu-affinity-lane. ## 3.0.2.11 — 2026-08-10 — cpu-affinity-lane **Goal:** Reduce post-x9 local **41** harvest misses (esp. pos5) and protect DyzineFrame render FPS by giving each stack a preferred CPU on 4-core Pi 5. **Phase A (process pin):** - `cpu_layout.sh`: OS=0 NET=1 DF=2 EG=3 (override `OHD_CPU_*`; off: `OHD_CPU_ENABLE=0`). - `uartd_common.sh` + `hat_bridge.sh`: `taskset -c $OHD_CPU_EG` for uartd + bridge Python. - systemd drop-ins via `install_cpu_affinity.sh`: `CPUAffinity=` on `ethergold` / `ethergold-collector` (wired from `install_prod_service.sh` / collector / commission). - **`sync/update.sh`** runs `install_cpu_affinity.sh` after each tip update (same as NM eth unmanaged). Opt out: `SKIP_CPU_AFFINITY=1`. **Phase B (low risk IRQ/RPS):** - `apply_irq_affinity.sh` + oneshot `ohd-irq-affinity.service`: USB xHCI etc. → EG core; eth/wlan prefer NET core; RPS on eth/gretap/wg/wlan → NET bitmask. Best-effort only. Also applied from `update.sh` via `install_cpu_affinity.sh`. Includes **3.0.2.10** copper-x1-before-fanout. ## 3.0.2.10 — 2026-08-10 — copper-x1-before-fanout **Goal (all10 jpot soak `0016` `$1107.80` 37BC):** Master did not reseed progressive after a bank jackpot because hub treated **eth_rx of spoke 41** as “Master has the close.” Multiport progressive fanout + `non_local_eth_x1_fanout` exclusive clear ran **before** (and without) hub copper inject of winner 41 → Master re-TX elevated progressive; next JP paid **$1114.20** (non-reset). Product intent was always: **winner unicast progressive → winner x1 on Master copper → then multiport peer fanout.** **Fix:** 1. eth_rx of armed jpot 41 only stamps `bank_jpot_x1_eth_seen` + last_frame/queue time (`_note_bank_jpot_eth_x1_seen`). **Does not** multiport-flood or clear exclusive. 2. Multiport progressive flood only from copper seal: `_bank_jpot_x1_close_ok` (`inject_echo_ok` / `copper_packet_complete` / `inject_glued_0d`) → `_release_bank_jpot_eth_fanout`. 3. Retired living eth-only clear (`non_local_eth_x1_fanout` close_ok path is a no-op). 4. Soft peer-game open requires `fanout_released` (post-copper), not eth_seen alone. Includes **3.0.2.9** prog-csv-snapshot meter logging. ## 3.0.2.9 — 2026-08-09 — prog-csv-snapshot **Goal (all9 jpot soak `2113`):** progressive non-reset was only visible via climbing x9 `progwin_cents` (e.g. D249 158980 → 177140). `progressive_0a_csv_update` logged only game1/game16, so other meter slots and frozen reseeds were hard to judge from events alone. 1. **`progressive_0a_csv_update`:** emit full **game1..game16** pennies whenever the 0A set actually changes and CSV is rewritten. 2. **`progressives_csv_snapshot`:** once per ~60s from `capture_tick` (last 0A or on-disk CSV) with the same full game fields — stable trail even when values are frozen. 3. **`bank_jpot_award_x9_ok`:** include **`progwin_cents`** from the grant frame. Includes **3.0.2.8** spoke-key-perm. ## 3.0.2.8 — 2026-08-09 — spoke-key-perm **Goal (pos5 soak `1953`):** initial spoke EDGE connect often burned ~1–2 min on `spoke_enroll_error` / pipe not ready when `/etc/ethergold/spoke/wg_private.key` was root-owned. Directory probes could still look “writable” while the key file itself denied write → enroll loop → MRT if glass played early. **Fix:** `_keys_dir_ready_for_spoke` / `ensure_spoke_keypair` skip a tree that holds non-RW key residue, then mint or use keys under `~/.local/share/ethergold/spoke` (or configured `EDGE_SPOKE_KEYS_DIR`). No longer dies on EACCES against `/etc`. Includes **3.0.2.7** concurrent eth queue work. ## 3.0.2.7 — 2026-08-09 — concurrent-eth-queue **Goal (all10 soak `1827`):** multi-seat “retries” were mostly eth/txn queue pressure, not soft pad or inject `attempt>1`. 1. **EDGE_TOPOLOGY=spoke:** skip eth→UART inject of **peer** gameplay (request/response/ack/0D). Spokes only need *local* TID game legs on copper; peer frames were stacking `transaction_defer` → `queue_cap` abandon (~1k on pos1/4). Progressive 0A / STOM / bank-jpot winner path unchanged. 2. **queue_cap eviction:** drop lowest-priority deferred first (foreign game before local x9/x1/0D; bank jpot last). Cap **64→128**. 3. **master_phase_diag hub:** stamp Phase A from **eth_rx x0** when copper x0 is not local (hub inject path) so live A is usable under hub–spoke. C string only rev-aligned (no hot-path C change vs 3.0.2.6). ## 3.0.2.6 — 2026-08-09 — soft-pad-master-phase **Goal:** shave residual inject soft pad; **prove** Master Phase A on every hand. **Soft pads (C):** 1. **pre_bulk SLV-IN:** `slv_in_phase_active` used GET_REG (`pog422_slv_in_live`) every inject — multi-ms USB. Prefer GPIO1 event cache (`slv_in_active`). 2. **pre_tx=0 path:** no `inject_poll_trace` / capture poll on zero sleep. 3. **HW_DEBUG copper:** PRE_DE / POST_OFF copper GET_REG only with `POG422_INJECT_FORENSICS=1` (not bare HW_DEBUG). 4. **DONE earlier:** after peel harvest, send `INJECT_DONE` *before* HS arm / post-x9 grace GPIO polls (was post-x1 soft work before Python unlocked). **Master phase (Python):** emit `master_phase_diag`: - `hub_40_sw` — eth→UART 40 inject wall (rpc, pre_bulk, job, air, de_hold) - `copper_A_40_to_49` — copper gap + `master_beyond_hub40_us` (A minus hub inject) - `copper_B_49_to_41` — local phase gap + full 40→41 Pass soak: `pre_bulk_us` p50 ≪ 1–2 ms; `master_beyond_hub40_us` shows where Master sits beyond hub inject (~30 ms copper A still dominated by non-hub time). ## 3.0.2.5 — 2026-08-09 — inject-hotpath-peel-fix **Soak `1549` (EG 3.0.2.4, hub3+spoke4):** functional OK (48:48 hands), no jpot. But **no LNT win** — two regressions vs `0223`: 1. **`pre_bulk_us` still p50≈5988 µs** on every inject. Bench builds with `HW_DEBUG=1` forced `inject_forensics_full()` true → same pre-TX USB GET_REG pad. Copper timeline remains HW_DEBUG-gated separately. 2. **Spoke x9 peel worse:** `peel_first_byte_us` p50=**20063** (= full 20 ms first-byte gate); `uartd_post_burst` p50 **23.5 ms** vs **17.5 ms** on 0223. Cause: peel idle used `poll(POLLIN)` on the tty, but **dttty/wide ring RX does not wake POLLIN** — loop slept the full quiet window then read. **.3.0.2.5:** forensics_full = env only (not HW_DEBUG); peel idle back to short nanosleep (50 µs) + pull. Keep peel marks + cache preciseflags from .4. ## 3.0.2.4 — 2026-08-09 — inject-hotpath-peel **Soak `0223` residual:** every eth→UART inject paid **`pre_bulk_us` p50 ≈ 6 ms** with `idle_wait=0` — open-loop USB control forensics *before* DE (GET_RX_STATS, RX FIFO count, GPIO_MODE, UART_FLOW). Spoke x9 peel residual (~17 ms post-burst) is mostly Master wire RTT; first-byte/hard caps unchanged (.53 risk). **.3.0.2.4:** 1. **inject_burst hot path:** skip pre-TX FIFO/regs probes (and post RX-FIFO sample) unless `POG422_INJECT_FORENSICS=1` or HW_DEBUG. Cache preciseflags after first confirm. Seed GPIO1 edge trace from the `require_idle` SLV-IN sample (no second probe). 2. **Peel:** `poll(POLLIN)` instead of 50 µs nanosleep spin; early-exit `post_de` when local x1 frame complete; keep 20 ms first-byte / 45 ms hard. 3. **IPC:** `INJECT_DONE` peel trailer (`peel_first_byte_us`, `peel_hit_us`, `peel_src`, `peel_x1_msg`) for soak marks; existing `X1_ACK_FRAME` still fires pre-DONE (event ACK for Master eth-forward — not a new open-loop). Expect hub 40/41 inject wall **~−5–6 ms**; peel hit walls unchanged until Master responds earlier. ## 3.0.2.3 — 2026-08-09 — jpot-hub-seal-fanout **Field (soak `0202`):** after bank jpot on hub+spoke, spoke sealed clean but **hub** held bank_jpot x1 priority for **30 s** (`close_miss hold_timeout`), **zero** peer-40 copper injects while EDGE still received eth 40, and multiport fanout blasted **25** EDGE ports (`eth_tx 49`×26 + heavy `0D`). LNT “comms dead then horrible”. **.3.0.2.3 fixes:** 1. **Award seal on hub:** stop discarding `award_x9_ok` on first arm; stamp `arm_progressive` from the progressive frame (result_x9 often lands ms before arm; discard → `seal_hold no_award_x9`). 2. **Non-local hold exit:** after eth x1 + multiport fanout release, hub (winner TID ≠ local_slave) clears exclusive (`non_local_eth_x1_fanout`) — copper seal remains for **local winner glass** only. 3. **Soft-open peer injects:** once every armed jpot has eth_x1 + fanout, `_should_defer` / hub allow resume peer x0/x9 (even if local seal pending). 4. **Fanout blast:** peer ports = TID-mapped spokes first, else carrier-up; skip multiport if no peers (winner already unicast). No more full EDGE map flood. Known-good floor: **`3.0.2.1`** scoop; tip includes `.2.2` async AC. ## 3.0.2.2 — 2026-08-09 — inject-async-auto-close **Fix:** spoke x9 `inject_rpc_scoop` ~21 ms was Python blocking under `_tx_lock` for post-DONE **AUTO_CLOSE_DONE** (C still TX synth 0D). Soak `0137` (3.0.2.1): scoop p50 21 ms on msg=49 while hub 40 scoop already ~0.1 ms. **.3.0.2.2:** 1. After `INJECT_DONE`, still nonblock luxury scoop (`.2.1`); **no wait** for `AUTO_CLOSE_DONE` on success path (0D after DONE). 2. Miss AC still binds on inject result (uartd emits pre-DONE). 3. Success AC via capture/`on_auto_close_done` + FIFO `_ac_arm_q`. 4. `AUTO_CLOSE_DONE` promotes to capture queue (not discardable inject orphans). 5. Pending AC: no empty-echo reinject. Expect spoke msg=49 `inject_rpc_scoop_us` ≪ 5 ms; full eth_rx49→done ~−20 ms. Rollback / known-good floor: **`3.0.2.1`** inject-noblock-scoop. ## 3.0.2.1 — 2026-08-09 — inject-noblock-scoop **Fix:** inject RPC wall ~110 ms was ~25 ms DONE + **~83 ms post-DONE scoop** waiting on quiet-socket `settimeout(0.05)` (session `0121` inject-rpc-marks). **.3.0.2.1:** 1. After `INJECT_DONE`, drain ECHO/FORENSICS/COPPER/capture with **timeout=0** until empty (uartd already emits luxury immediately after DONE). 2. Optional ≤5 ms arm only if forensics still missing (race). 3. Auto-close: targeted wait for `AUTO_CLOSE_DONE` only (5 ms slices), not a blanket 50–120 ms quiet scoop. Expect hub peer-**40** `gpio_on→stage` / `inject_rpc_scoop_us` ≪ 10 ms typical. Master copper RTT (~110 ms) unchanged. Rollback: `3.0.1.58` inject-rpc-marks. ## 3.0.1.58 — 2026-08-09 — inject-rpc-marks **Diag only:** after `.57` proved idle_wait=0 and C burst ~20 ms while Python wall ~112 ms, add marks that split **uartd pre-burst / burst / post-burst / job→DONE** from **Python INJECT→DONE** and luxury scoop. **.58 fields on `inject_stage_diag`:** | Field | Source | |-------|--------| | `uartd_pre_burst_us` | handle_inject enter → `inject_burst` | | `uartd_burst_wall_us` | `inject_burst` wall | | `uartd_post_burst_us` | burst return → `INJECT_DONE` send | | `uartd_job_to_done_us` | full worker until DONE | | `inject_rpc_wait_us` | Python INJECT send → DONE recv | | `inject_rpc_scoop_us` | DONE → client return (echo/forensics) | | `inject_capture_ipc_n` | capture IPC drained while waiting | | `rpc_minus_uartd_us` | Python wait beyond uartd job→DONE | | `uartd_outside_burst_us` | uartd time not in inject_burst | Rollback: `3.0.1.57` inject-stage-forensics. ## 3.0.1.57 — 2026-08-09 — inject-stage-forensics **Diag only (no TX policy change):** hub eth→UART gameplay 0x40 shows ~113 ms wall while DE hold ~11 ms — stage split instrumentation to prove pre-DE pad (idle wait). **.57:** 1. C inject measures `idle_wait_elapsed_us`, `pre_bulk_us`, `bulk_wall_us`, `require_idle`, `idle_wait_budget_ms`, `idle_saw_busy` on every burst. 2. Fields on `INJECT_DONE` (trailer; short DONE still ok). 3. Bridge emits `inject_stage_diag` (events.jsonl) with `pre_de_pad_us` (= total − de_hold) and `pre_bulk_minus_idle_us`. **Short soak focus:** hub + one spoke; confirm version/rebuild uartd; greps: `inject_stage_diag` on msg=40. Rollback: `3.0.1.56` jpot-lift-no-thrash. ## 3.0.1.56 — 2026-08-08 — jpot-lift-no-thrash **Field:** after bank jackpot, console flooded with `bank_jpot_x1_priority_lift` every main-loop poll. **Cause:** `_lift_bank_jpot_x1_from_transaction_deferred` promoted foreign (and other currently-deferrable) gameplay x1/0D out of Queue B; under local exclusive they were immediately re-stashed → permanent lift/stashed thrash + log spam through the 30 s arm / exclusive window. **.56:** 1. Only lift frames that may inject **now** (skip if `_should_defer_transaction_inject` would bounce them). 2. Rate-limit `bank_jpot_x1_priority_lift` log (same cadence as inject_defer). 3. Keep `.55` hand-open-on-miss policy. Rollback: `3.0.1.55` hand-open-on-miss. ## 3.0.1.55 — 2026-08-08 — hand-open-on-miss **Policy:** empty x1 harvest = **hand still open**. Do not invent 0D / x1; let Master re-49 and a real T340+ **41** complete the path (soak 1941: Python 0D after miss then `x9_eth_inject_dedupe` / `seq_already_closed` killed glass). **.55:** 1. **`auto_close_miss`:** never arm Python deal 0D (CGS or not). Emit `status_0d_auto_close_miss_hand_open`; clear Keno eth-inject age for that TID+seq. 2. **`mrt_x1_infer_*` disabled:** empty wait no longer `_mark_local_x1_pending(infer)`. 3. **While `await_x1` and not closed:** soft-dedupe does **not** drop Master re-49. 4. **Same-seq Master re-49** while still await_x1: refresh await clock + note `x9_retry_seen` (keep original peel path armed for second inject). 5. Keep peel timings of `.54` (no 20 ms hard-hunt); keep exclusive + DONE-before-0D. Rollback: `3.0.1.54` restore-x1-peel. ## 3.0.1.54 — 2026-08-08 — restore-x1-peel **Field:** `.53` (`x1-hunt-20ms`) → **continuous** glass MRT. Hard 20 ms peels + no HS/post_de stack + 20 ms pending → mass `NO_X1` → Python 0D before real 41 / Master retry can complete. **.54 reverts .53 hunt policy; keeps .51 exclusive + .52 DONE-before-0D:** 1. Imm peel hard wall **45 ms** / first-byte **20 ms** again. 2. POST_DE 25 ms + secondary 15 ms; HS wait **40 ms**; pending **100 ms**. 3. Auto-close again stacks imm → post_de → HS on miss. 4. Python local FIFO grace **25 ms**; late_0D **150 ms**; x1-infer **25 ms**; AC scoop restored (0.12/0.20 s). 5. Synth 0D idle **25 ms** (.52). Rollback: `3.0.1.52` if exclusive path also suspected (not the .53 regression). ## 3.0.1.53 — 2026-08-08 — x1-hunt-20ms **FAILED IN FIELD** — continuous MRT. See **3.0.1.54**. **Policy (rejected):** post-x9 **x1 insurance ≤ 20 ms.** If we miss x1, Master re-sends 49 — do not invent 100+ ms stacked peels. **.53 (reverted):** 1. C: imm peel hard wall **45 → 20 ms**; quiet activity gate **20 → 10 ms**. 2. C: POST_DE / secondary / HS wait / AUTO_CLOSE_PENDING / post-x9 grace **≤20 ms**. 3. C auto-close: **do not stack** full post_de + HS wait after imm peel. 4. C: synth 0D idle wait **≤20 ms**. 5. Python: local grace / late_0D / x1-infer **≤20 ms**; scoop shortened. Rollback: `3.0.1.54` (or `3.0.1.52`). ## 3.0.1.52 — 2026-08-08 — x9-done-before-0d **Soak `2026-08-08_1754-UTC_pos5-mrt-eg51-soak`:** every local 49 inject held Python `inject_ioctl` **~114 ms** (p50) while wire TX of 49 was only **~17 ms**. Eth **41** already arrived **~40 ms** into that window; remaining time was **auto_close synth 0D** with **150 ms** HS-idle wait *before* INJECT_DONE. **.52:** 1. After x9 TX + x1 peel: send **INJECT_DONE** (and luxury echo/forensics) **before** synth 0D — unlock eth→UART for concurrent traffic. 2. X1_ACK IPC still fires at harvest (before DONE) for Master eth-forward. 3. Synth 0D runs post-DONE with `AUTO_CLOSE_0D_IDLE_WAIT_MS` **150 → 25**. 4. AUTO_CLOSE_DONE (ok / fail) after 0D; Python scoop extended when auto_close armed. 5. Keep `.51` exclusive fence; no longer rely on long miss-defer. Rollback: `3.0.1.51` local-game-txn-exclusive. ## 3.0.1.51 — 2026-08-08 — local-game-txn-exclusive **Policy (operator):** once a *local slave* x0 (gameplay request) starts, treat the hand as an exclusive **local game transaction** until 0D. Keep other eth→UART traffic out — no “under load” excuse for barging progressive / foreign / stacked same-seq 49 into the inject path. Stretching NO_X1 Python defer does not fix glass MRT when CGS 40 stole the harvest window (soak 1634). **.51:** 1. **`_local_game_txn_exclusive()`** from local trunk open (x0) through await_x1 / await_0d / synth 0D (spokes only; hub Master stays narrow 0D fence so peers do not freeze). 2. While exclusive: defer **all** eth→UART except this seat’s x9/x1/0D/skill, STOM, bank-jpot priority. Progressive and foreign always wait (`transaction_defer reason=local_game_txn_exclusive`). 3. **Second same-seq x9** after first inject: peel/hold (extends .17) so stacked eth 49 during first DE cannot land (1634). 4. **CGS during miss:** peer gameplay-request in inject RX → **no Python 0D** (`status_0d_auto_close_miss_cgs_no_synth`); do not “wait longer then close”. 5. **Revert** Python miss defer to **20 ms** (late AUTO_CLOSE only). Drop .50 long window as reliability strategy. 6. Events: `local_game_txn_open` / `local_game_txn_close`; local trunk stale bound `LOCAL_GAME_TXN_MAX_MS` (2 s). Rollback: `3.0.1.49` if exclusive starves multi-TID on a mis-labeled hub seat. ## 3.0.1.50 — 2026-08-08 — late-x1-miss-defer **Soak `2026-08-08_1634-UTC_pos5-mrt-eg49-soak`:** pos5 `.49` autoplay ~76 good closes then glass MRT. **Not** reopen/late_0d: one `auto_close_miss` harvest=0 with concurrent CGS `40` during 49 inject; Python 0D closed ~100 ms later; true local `41` arrived **~212 ms** after miss — already `seq_already_closed`. Autoplay issued no further 40. Pos1/2/4 zero miss on same bank. **.50:** 1. `STATUS_0D_PYTHON_FALLBACK_DEFER_MS` **20 → 250** ms (late x1 can win). 2. Concurrent gameplay-request in inject peer RX → `STATUS_0D_PYTHON_FALLBACK_DEFER_CGS_MS` **350** ms, reason `auto_close_miss_cgs`. 3. Late local x1 while miss-defer armed: clear ready timer, emit `status_0d_python_fallback_late_x1`, schedule close with x1 (not pure NO_X1). Keep .49 recovery/ok reopen. Rollback: 3.0.1.49 if miss defer starves true misses into glass MRT without close. ## 3.0.1.49 — 2026-08-08 — reopen-recovery-ok **Field:** pos5 continuous MRT again after `.48` settle-only reopen (2500 ms on *all* closes). Glass re-Play after miss/close or product MRT stayed `seq_already_closed` until long settle — same latch class as soak 1436. **.49:** 1. Close quality **`ok` | `recovery`**: `auto_close_miss` Python fallback and `mrt_risk_late_0d_after_x1` stamp **recovery** → same-seq local wager reopens **immediately**. 2. Normal uartd good close stamps **ok** → reopen only after `MRT_HAND_REOPEN_MIN_AGE_MS` (**400 ms**, was 2500) — still blocks 1513 CGS ~1 ms re-wager → false late_0d. 3. Events: reopen `reason=same_seq_local_wager_recovery|settled`, quality field; skip still `too_soon_after_close` for ok path. Rollback: `3.0.1.48` only if CGS-class late_0d returns under 400 ms settle. ## 3.0.1.48 — 2026-08-08 — reopen-settled-only **Soak `2026-08-08_1513-UTC_pos5-mrt-after-eg347`:** on **.47**, pos5 advanced play_sequences and still hit product MRT. Metrics: **8×** `mrt_risk_hand_reopen` paired with **8×** `mrt_risk_late_0d_after_x1` (bound_ms=150). Pattern: good uartd close → CGS same-seq re-wager ~1 ms later → reopen cleared close-done → re-armed `await_0d` without a second 0D → classic late_0d. Continuous `seq_already_closed` latch from 1436 is gone; reopen was too eager. **.48:** 1. **`_mrt_risk_reopen_hand_on_local_wager` gated:** only after close-done age ≥ `MRT_HAND_REOPEN_MIN_AGE_MS` (2500 ms) and not while open close-wait. Skips emit `mrt_risk_hand_reopen_skip` (`too_soon_after_close` / `open_close_wait`). Settle window still recovers 1436-class latch without reboot (glass clear + Play is multi-second). 2. Stamp `_mrt_risk_close_done_ns` with `_mrt_risk_mark_seq_closed`. Also residual this soak (not fixed here): **2×** `status_0d_auto_close_miss` (NO_X1); Master 40→49 RTT outliers to ~737 ms. Rollback: `3.0.1.47` (only if continuous latch returns and skip age is wrong). ## 3.0.1.47 — 2026-08-08 — reopen-same-seq-close **Soak `2026-08-08_1436-UTC_pos5-immediate-mrt`:** continuous glass MRT while EG logged `seq_already_closed` / `already_closed` on every retry. Only one NO_X1 seed; glass stayed on play_sequence **2** forever; bridge reboot cleared the close-done map and restored play. Pos1–4 healthy on same tip. **.47:** 1. **`_mrt_risk_reopen_hand_on_local_wager`:** local same-seq gameplay request (`0x40` etc.) after that seq was marked closed → clear `_mrt_risk_close_done_seq` + local close-wait / synth_sent → emit `mrt_risk_hand_reopen`. Next eth→UART x9 can arm auto-close again (no Pi reboot required). 2. **Spoke WG keys:** `ensure_spoke_keys_dir` / robust `ensure_spoke_keypair` fall through from unreadable/unwritable `/etc/ethergold/spoke` to `~/.local/share/ethergold/spoke` (fixes `Permission denied: wg_private.key` enroll spam). Same pattern as hub peer store (3.0.1.30). Rollback: `3.0.1.46`. ## 3.0.1.46 — 2026-08-07 — revert-deal-open-guard **Soak `2026-08-08_0011-UTC_master-no-slave-cgs-eg345` (pos1–4, tip .45):** Master T340+ saw **no slave data**; slaves stuck in **CGS**. Master eth_rx many `0x40` but `transaction_defer` (game_transaction_gate + queue_cap) starved UART inject; almost **no copper `packet_complete` 40** on Master. Root cause: .45 “serialize foreign 0x40 while any trunk open” — first open trunk never cleared fast enough → peer wagers deferred forever/to cap → glass never saw wagers. **.46 REVERT of .45 inject gates:** 1. Remove `_deal_open_guard_active` and same-TID re-wager defer under await_x9. 2. Remove serialize-foreign-0x40-while-any-trunk-open. 3. Restore pre-.45: gameplay requests (same-TID supersede + other TID 40) pass the transaction gate; soft `0x0A` still defers while trunk open. `MASTER_DEAL_OPEN_GUARD_MS` remains in core as docs only (not wired). MRT late first-deal (2258) is still open — do **not** re-land serialize without a proven flush path. Rollback: `3.0.1.44` (skip .45); operational tip is this rev after .45 field fail. ## 3.0.1.45 — 2026-08-07 — master-deal-open-guard ⚠️ FIELD FAIL **MRT soak 2026-08-07_2258 (pos5 D249):** Master eth_rx40→uart49 ~276 ms on the miss hand vs ~113 ms healthy. Inject queue wait for first D249 40 was ~0.4 ms — not eth backlog. Same-seq CGS re-wager ~170 ms after first 40 was **re-injected onto Master copper ~18 ms before a healthy first-deal 49 window**; 49 landed ~104 ms after the **second** 40 only. Glass had no x1 for the late first result. **.45 Master deal-open guard (BAD — do not ship):** 1. **`MASTER_DEAL_OPEN_GUARD_MS` (140 ms):** while trunk open for a TID and x9 not yet marked, **defer same-TID gameplay re-wager eth→UART inject**. 2. **Serialize foreign new 0x40:** while any trunk is open, defer **other TID** gameplay requests. 3. x1 / 0D / STOM / P0 still never deferred. **Field fail soak 2026-08-08_0011:** Master glass dark / slaves CGS — see **3.0.1.46**. Rollback: `3.0.1.44` or deploy **3.0.1.46**. ## 3.0.1.44 — 2026-08-07 — x1-rx-pending-100 **NO_X1 MRT at pos5 18:16:** hard-RT auto-close miss (post_de filled with next keno `0x40`); instant Python P0 raced C pending; peel exited early on complete non-x1 frame. **.44:** 1. **C `AUTO_CLOSE_PENDING_NS` 200 ms → 100 ms** (glass timeout territory). 2. **Imm x1 peel:** after DE→RX, hunt local ACK to hard_cap — do not stop on complete non-x1 frames; trust RxD over “quiet HS” frame-exit. 3. **Python miss fallback** defers **`STATUS_0D_PYTHON_FALLBACK_DEFER_MS` (20 ms)** for late `AUTO_CLOSE_DONE`; cancel if uartd closes first. Extra 0D is harmless to T340+; defer is for late C close, not multi-0D fear. Rollback: `3.0.1.43`. ## 3.0.1.43 — 2026-08-07 — 0d-single-owner **Soak 2026-08-07_1944 (pos5 MRT + dual 0D):** uartd auto-close succeeded (`status_0d_synth_done source=uartd_auto`), then Python re-armed after `_release_synth_sent` + `_queue_synth` bypass of the uartd gate and fired a second full-forensics 0D. Second owner is wrong; MRT `seq_already_closed` noise. **.43 single owner for deal close:** 1. **Python deal 0D never queues while uartd owns** — `_queue_synth` / `_try_synth_pending_for_round` / schedule all gate on `_python_may_deal_synth_close` (fallback only). 2. **After auto_done (synth_sent):** re-harvest of local x1 → `status_0d_synth_skip reason=already_closed` — no re-pending, no release of synth_sent that cleared the dual path. 3. **auto_close_miss only:** arm one Python P0 via `_arm_python_deal_close_fallback` + `_synth_retry_rks`. 4. Skill-grant close remains Python (C auto-close is deal path only). Rollback: `3.0.1.42`. ## 3.0.1.42 — 2026-08-07 — jpot-seal-reopen-hand **Unclaimed progressive JP14 soak (`logs` 2026-08-07 T15:12):** copper grant + fireworks fired; winner T340+ never “noted” (seq closed by empty `41`); Master sealed-hold on **pre-arm empty x1 inject_echo** twice; arm **hold_timeout** + STOM starve. **.42 fixes (seal + hand reopen, keep .38 anti-steal):** 1. **Bank progressive arm reopens hand** for winner TID/seq: clear `mrt_risk_close_done_seq` + empty-close block so auto-close / MRT are not stuck `seq_already_closed`. 2. **Seal requires `award_x9_ok` AND post-arm x1** (not post-arm alone). 3. **Reinject stamps** `_bank_jpot_x1_eth_queued_ns = now`; only reinject when a **post-arm** last_frame exists (purge / refuse pre-arm empty as last_frame). 4. **Local progressive issue** injects front-of-queue at bank-jpot class while armed. 5. While bank-jpot arm (or award set): **native Master eth `0D` not blocked** by `status_0d_block_all` (synth remains backup). 6. Hold-timeout path still clears priority (no flood) so STOM recovers. Rollback: `3.0.1.41`. ## 3.0.1.41 — 2026-08-07 — pogcc-telemetry-diag When `pogcc` is null, machine-status includes `pogcc_telemetry_path`, `pogcc_telemetry_present`, and `pogcc_hint`. status/ created 775+owned by topbox so DyzineFrame can write pogcc-telemetry.json when bridge created the dir as root. Rollback: `3.0.1.40`. ## 3.0.1.40 — 2026-08-07 — machine-status-json **pog-atlas machine status file:** durable `~/EtherGold/status/machine-status.json` (outside tip trees — survives `update.sh` / code rsync). EtherGold merges: - `status/pogcc-telemetry.json` — DyzineFrame from POG-CC `/telemetry` (~5s): four INA rails V/I/P, upper/lower temps °F, credit balance, game menu - `progressives.csv` — all 1..16 pennies + game names - `status/last-progressive-hit.json` — last bank progressive grant (UTC + game_index/name + progwin) Refresh on 0x0A CSV updates, progressive hit, and bridged packet traffic ~5s. Not under `EtherGold/sync/` package paths. Rollback: `3.0.1.39`. ## 3.0.1.39 — 2026-08-07 — heltec-boot-arm-force-cg **Gap-fill without replacing SDS:** after Heltec USB-up, DyzineFrame writes `~/EtherGold/pogcc-boot-arm.json` from POG-CC `/hello///`. Bridge always runs current-game observe; continuous write still gated by `ETHERGOLD_CURRENT_GAME_IPC`. When arm is present, the **next local-TID** mapped gameplay **force-publishes** `current-game.json` even if index unchanged (so DF inotify re-applies SuperBall after Heltec flash). Arm file cleared on publish. SDS enter-game remains primary when live. Rollback: `3.0.1.38`. ## 3.0.1.38 — 2026-08-06 — jpot-award-before-seal **Stolen progressive (JP3 soak `2026-08-06_2040`):** Master sealed bank progressive and reset game11 meters by eth→UART-injecting a **stale empty `41`** that was already queued for the winner TID ~200 ms *before* the progressive grant (`progwin>0`). Glass never showed hits / Jackpot Won; CGS followed. - On first bank-jpot **arm**: purge pending/deferred/spoke **pre-arm x1** for that TID; stamp `arm_ns`; clear empty-close block. - **Seal** (`inject_echo_ok` / copper) only for **post-arm** x1 evidence (`bank_jpot_seal_hold` if pre-arm/stale). - Winner only: eth→UART inject of progressive grant; peers skip (`bank_jpot_peer_no_uart_inject`) — OHD via bank_wlan remains Master-side. - Local **empty close** (x1 without recent result x9) blocks progressive award inject (`bank_jpot_empty_close_no_award`). - **hold_timeout / max_reinject**: drop stashed eth fanout (no multiport flood). Keeps 3.0.1.37 never soft-dedupe of local progressive grant. Rollback: `3.0.1.37`. ## 3.0.1.37 — 2026-08-06 — jpot-award-never-dedupe **Product law:** local-terminal bank progressive grant (`progwinval>0`) is never `x9_eth_inject_dedupe`'d. Award inject is not optional. Soak `2026-08-06_1744-UTC_pos1-5-jpot-reset-no-award`: cash SuperBall `0x49` (progwin=0) inject stamped TID+seq; true jpot `0x49` (progwin $1631.50) ~292 ms later hit **dedupe before bank-jpot arm** → pos4 no eth→UART award, while Master copper-sealed x1 and **0A game11 reset** ($1631→$1000). Mirror dealkiller: reset without Jackpot Won. - Local + `is_bank_progressive_jackpot_frame` → never soft-dedupe (do not require arm already open — eth order is dedupe then arm). - Peer progressive flood still age-dedupes; pending match ignores non-jpot cash for bank_jpot frames. - Fix `bank_jpot_eth_fanout_release` TypeError: `router.port_names` is a method (`port_names()`) — was `'method' object is not iterable`. Rollback: `3.0.1.36`. ## 3.0.1.36 — 2026-08-06 — jpot-seal-abort-window `BANK_JPOT_X1_SEAL_MS` **250 ms** (was 10 000). Master progressive grant aborts in ~**200 ms**; post-copper sealed reinject must cover only that race (2nd/3rd Master `49` re-TX), not linger for seconds. Real path is still exclusive until copper so the first x1 hits UART before / with the first retry. Rollback: `3.0.1.35`. ## 3.0.1.35 — 2026-08-06 — jpot-copper-seal-reset **Dealkiller:** Slave shows progressive win while Master 0A meter keeps climbing (casino false advertising). Soak 2026-08-06_1651 SuperBall TID 37BC `progwin=$1572.27`: Master **did** get eth x1 at **+48 ms**, but copper inject echo only at **+172 ms** — *after* Master already re-TX progressive `49` at **+163 ms**. EG 3.0.1.34 cleared exclusive arm on `eth_rx_x1`, so Master progressive retries #2/#3 got **no** x1 reinject; arm #3 `hold_timeout` 30 s; **game11 never reset** on copper 0A (OHD stayed $1572.xx). Fix: - **Copper seal only:** exclusive clear + `bank_jpot_x1_sealed` on `inject_echo_ok` / copper PC / glued 0D — **not** eth_rx. - eth_rx still releases multiport fanout (`bank_jpot_x1_eth_seen`) for CGS. - Retain last x1 frame for `BANK_JPOT_X1_SEAL_MS` after copper only (see 3.0.1.36 for ms) so Master progressive re-TX can reinject without a second 30 s arm (`bank_jpot_x1_sealed_retry` / `master_progressive_retry_sealed`). Rollback: `3.0.1.34`. ## 3.0.1.34 — 2026-08-06 — jpot-clear-on-x1-evidence Master (and any arming seat) bank-jpot exclusive arm cleared when winner x1 is **proven**, not only when copper `packet_complete` sees that x1 after inject. Soak `2026-08-06_1258-UTC_pos1-5-jpot-hold-timeout-eval`: every Master arm hit `hold_timeout` @ 30 s while winner seats `close_ok` in ~0.12 s. Master **did** `eth_rx` + inject_echo **match** the x1 (`bank_jpot_x1_inject_done`); clear used to require copper PC of peeled self-echo (never fires) → 30 s x0 inject defer / CGS. - Clear evidence: `eth_rx` of armed TID x1; inject echo `exact|prefix|embedded` or non-empty echo; copper PC (unchanged); glued Master `0D` after peel. - `bank_jpot_x1_close_ok` gains `evidence=…`; clear reason matches. - `bank_jpot_x1_bus_trail` when inject RX has peer payload after self-echo (expect glued Master 0D <3 ms when present — soak often had exact 12 B only). Keeps 3.0.1.33 never-defer close legs. Rollback: `3.0.1.33`. ## 3.0.1.33 — 2026-08-05 — jpot-never-defer-close **Policy:** never transaction-defer gameplay **x1** or **0x0D** (incl. `status_0d_synth`) for any TID while bank progressive close is armed. EG 3.0.1.27 exclusive path deferred *everything* except winner x1 + progressive issue, so a non-winning slave's own close legs waited behind the jackpot window → true MRT (`x1` without timely `0x0D`) and post-jpot CGS for 1–2 plays even when jackpot meters reset. - `is_gameplay_close_leg`: any game ACK or status 0x0D / synth. - `_is_bank_jpot_close_path_frame` / cancel / lift / hub allow: close legs pass. - Soft progressive / peer x0 / non-jpot x9 may still yield during the arm window. - Winner x1 remains highest bank-jpot priority; eth no-repeat (3.0.1.32) kept. Rollback: `3.0.1.32` (jpot-eth-no-repeat). ## 3.0.1.32 — 2026-08-05 — jpot-eth-no-repeat Fix: after bank progressive, multiport eth delivery was re-playing the **same progwinval** on seats (and re-arming 30s exclusive close on every spoke) → operator sees the same jackpot over and over on every game path. - **eth_rx arm** bank-jpot x1 priority only when the **winner TID is local** (peers still inject progressive for bank bells; no 30s freeze-as-winner). - **Post-x1 eth fanout** excludes the winner port (unicast hold already delivered the grant; no second inject + re-arm). - **Eth inject dedupe** for bank-progressive re-floods (non-winner and post- clear winner); local winner **while still armed** still never-dedupes so Master progressive retries can re-inject for close. Keeps 3.0.1.31 winner-unicast-until-x1. Rollback: `3.0.1.31` (jpot-defer-fanout) or prod-safe `3.0.1.30` if defer-fanout still suspect. ## 3.0.1.31 — 2026-08-05 — jpot-defer-fanout Bank progressive eth: **winner unicast first**, full multiport fanout only after Master has the winner **x1**. Avoid out-of-order bus clog that stalls other seats' SuperBall (and similar) **game requests** (operator "Contacting Game Server" / multi- round abort) while progressive close is still open. - Master `uart→eth` / edge relay of `progwinval>0` x9 (or skill grant): `only_ports` = winner TID gretap until release; stashed for deferred flood. - **Release** on Master `eth_rx` of that TID's x1, copper `packet_complete` x1, or bank-jpot arm clear/miss (`bank_jpot_eth_fanout_release` / `_hold` events). - SCR (Send Com Request Local Net Tests) bank flood path unchanged (not CGS). Rollback: `3.0.1.30` (auto-edge-cgs-path). ## 3.0.1.30 — 2026-08-05 — auto-edge-cgs-path Zero manual ops for star CGS: enroll + Master path recover without EDGE_HUB_IP or chmod. - **Spoke uart→eth:** unmapped TID no longer black-holes local x0 (`only_ports= None` → all local eth ports = hub uplink). Hub still uses `[]` for unknown (no multi-spoke flood on map miss). - **Hub peer store:** `ensure_hub_state_dir` falls through to `~/.local/share/ethergold/hub` when `/etc/ethergold/hub` is not writable; merges any existing system `peers.json` for read. - **Spoke enroll:** poll bank-wlan `EDGE_UPLINK_IP_LEARNED` (and env/WG endpoint) for life of bridge — no one-shot `spoke_enroll_skip no_hub_ip`. - **Hub bank-wlan:** auto-publish `_guess_uplink_ip` as `hub_uplink_ip` so spokes learn without manual env. Rollback: `3.0.1.29` (spoke-tx-arbiter). ## 3.0.1.29 — 2026-08-05 — spoke-tx-arbiter Star hub inject + eth fanout: prioritize local gameplay with per-spoke buffers and per-TID transaction FSM; copper stays one-at-a-time. - **Eth:** TID is always on the wire; “unknown” means **TID→gretap map** missing. Unmapped gameplay never multi-port floods (`only_ports=[]`, log `tid_port_unknown`). Bank-wide only: progwin>0 x9, progressive **0A**, SCR. - **Per-spoke inject queues** (`_spoke_queues`) + copper arbiter rebuild of `_pending` by score (P0 / bank jpot / open-tx required legs / local terminal; soft 0A yields). - **TidTxFsm** per TID: `idle → await_x9 → await_x1 → await_0d → idle` (events `tid_tx_fsm`). Soft **0A** defers while any FSM open; 0A soft lane is latest-wins. - Bank jpot exclusive close path retained (3.0.1.27/28). Rollback: `3.0.1.28` (jpot-x1-copper-clear). ## 3.0.1.28 — 2026-08-04 — jpot-x1-copper-clear Bank progressive close stays armed until Master **copper** `packet_complete` of the winner x1 — not until eth→UART inject_done / self-echo alone. - **Clear only on** copper gameplay ack PC for armed TID (`bank_jpot_x1_close_ok` + `bank_jpot_x1_priority_clear` reason=`copper_packet_complete`). - **Do not clear** on `eth_to_uart_done` of bank-jpot x1; emit `bank_jpot_x1_inject_done` with echo_len / post_de_len / reinjects (self-echo alone never closes). - **Reinject:** empty inject echo → queue `bank_jpot_x1_reinject`; Master progressive re-arm with stored frame → reinject (cap `BANK_JPOT_X1_MAX_REINJECT=12`). - **Miss:** hold timeout or max reinjects → `bank_jpot_x1_close_miss` then clear (reason `hold_timeout` / `max_reinject`). - Arm payload adds `rearm`; clear payload includes forensics fields. Rollback: `3.0.1.27` (jpot-close-primary). ## 3.0.1.27 — 2026-08-04 — jpot-close-primary Bank progressive jackpot close is exclusive primary until the winner’s x1 injects — no IOP required. Closer to native: Master grant retries still get out; everything else stays out of the way. - **While bank jpot x1 armed:** eth→UART inject is only winner x1 + progressive-issue frames (`progwinval>0`). All other eth inject is transaction-deferred. Hub peer gameplay yields for the whole arm window. - **On arm:** cancel competing `_pending` into Queue B; lift armed x1; reorder. Events: `bank_jpot_x1_priority_cancel`, arm fields `cancelled_pending`. - **Master retries preserved:** never eth-inject-dedupe bank progressive frames; progressive issue bypasses await_x1 peel / post-x9 grace / inject-abort cool while close is armed. Do **not** starve Master 0x49 retries onto the seats. - **Close path speed:** bank-jpot x1 uses p0-style HS ready (skip post_hs settle). - **Fanout:** only bank progressive-win x9 full-floods for bells. Cash/other x9 with unknown TID→port no longer flood all hops (`only_ports=[]`). Fix: empty `only_ports` was treated as flood (falsy list). - Cash eth inject dedupe window while armed: **2000 ms** (non-jpot only). Rollback: `3.0.1.26` (superball-sheet11). ## 3.0.1.26 — 2026-08-04 — superball-sheet11 Phase 2 SuperBall cash uses **SGVI sheet 11** mult matrix (not TouchEasy sheet 10). Validated against bank help paytable at 1 credit / $0.25 (pos5 `T340_0002`). - TouchEasy (9309) unchanged — sheet 10. - SuperBall (930A): sheet 11 cash + last-ball **×4** when 20th ball is a pick hit. - BAJ (Buy-A-Jackpot) tables **not** evaluated; progressive remains Phase 1 via `progwinval` (`progwin_phase1` skip). Field BAJ option (e.g. table 10) only changes when the game routes progressive vs cash. Rollback: `3.0.1.25` (superball-keno-p2). ## 3.0.1.25 — 2026-08-04 — superball-keno-p2 Phase 2 cash early-win: **SuperBall Keno (930A)** in the allowlist (was `game_unsupported`). - Initially reused sheet-10 mults (corrected in 3.0.1.26 to sheet 11). - **Last draw ball** (20th ball @ offset 21+19) in the player's picks → cash win **×4**. - Progressive / Buy-A-Jackpot still **Phase 1** only (`progwin_phase1` skip when `progwinval != 0`). - Events: `early_win_keno_eval` / `_hit` include `last_ball_hit`, `superball_x4`, `base_pay_mult`, `base_win_cents`. Rollback: `3.0.1.24` (bank-jpot-x1-prio). ## 3.0.1.24 — 2026-08-04 — bank-jpot-x1-prio Inject priority class: after Master bank progressive jackpot (`progwin>0` x9 / skill grant), the winner seat’s gameplay **x1** eth→UART inject is **head-of-queue** for 30 s — ahead of lab synth / foreign `80` / STOM FIFO peers. - **Arm** on Master `uart_to_eth` and on eth_rx of bank jpot (`bank_jpot_x1_priority_arm`). - **Class** `INJECT_PRIORITY_BANK_JPOT_X1` (20) between STOM (10) and local P0 status-0D synth (30). - Never transaction-deferred; lifted from Queue B; grace holds bypassed like STOM. - Events: `bank_jpot_x1_priority_arm` / `_clear` / `_lift`; `eth_rx_queued` adds `bank_jpot_x1_priority` + `inject_priority`. Root-cause fix for star soak where hub `queue_wait_us≈429ms` delayed progressive `0x81` while Master waited (meter/reset failure). Rollback: `3.0.1.23` (hub-hot-enroll-eth). ## 3.0.1.23 — 2026-08-03 — hub-hot-enroll-eth Field hub/spoke: Master that is already running must learn new spokes **without restart**. - **`DirectionalEthRouter.add_port`**: hot-open AF_PACKET on `eg-sNN` after live enroll. - **Eth RX select**: refresh `router.sockets()` every poll (was cached at thread start — new gretaps never appeared in `select`). - **Enrollment hub**: `on_peer_ready` after accept / restore / reuse → `add_port` + `.active_spoke_gretaps` append. Prefer request `spoke_id` / `EDGE_SPOKE_ID` when free. - **Spoke auto-enroll**: topology `spoke` posts UDP enroll retries to hub Endpoint from `wg0.conf` (events `spoke_enroll_response` / `*_error`). Bridge runs as topbox: `/etc/wireguard` is often root-only — enroll must tolerate `PermissionError` (`EDGE_HUB_IP` env preferred) and never abort the UART session. - HMAC + open-window / allowlist rules unchanged. `peers.json` remains live roster. Rollback: `3.0.1.22` (lab-withhold-status-0d). ## 3.0.1.22 — 2026-08-02 — lab-withhold-status-0d Lab tool for intentional Slave MRT on a seat with POG-CC (pos5): withhold the local synth `0x0D` after eth→UART `x9` / local `x1` for a limited number of rounds so the T340+ times out (~125 ms). Correlate with Heltec `POG-CC SDS error pending` stamps (DF `2026-08-02-g+`). - **Python:** `--lab-withhold-status-0d` / `FIXTURE_LAB_WITHHOLD_STATUS_0D=1` (alias `LAB_WITHHOLD_STATUS_0D=1`) with `*_COUNT` (default **1**). Skips uartd `auto_close` arm and Python `status_0d_synth` for those rounds; Master eth `0x0D` still blocked by existing `status_0d_block_all`. Emits `lab_withhold_status_0d` / `lab_withhold_status_0d_exhausted`. Forbidden with `--prod`. - **Launcher:** `run_bridge_jun11_lab_withhold_0d.sh` (FIXTURE_POS=5). - uartd/inject tip string only (no C close-path change). Driver still **2.2.2.10**. Rollback: `3.0.1.21` (x1-ack-wire-len). ## 3.0.1.21 — 2026-08-02 — x1-ack-wire-len Pos2 SpinBall progressive unpaid on tip `3.0.1.20` (`2026-08-02_1515-UTC`, TID `38D7`, progwin **$2,824.56** @ 08:49 EDT): `extract_local_x1_frame` used first-EOP hunt, but **EOP is `0x0D`**, so a **6 B** false SOH..EOP with local TID was emitted as `IPC_X1_ACK_FRAME`. Python eth-forwarded junk (`peer_len=6`, `peer_msgs=null`), latched `early_x1_eth_done`, then **skipped** the real 12 B `0x81` in the post-inject peer path. Master saw `80`→`89` + bank WLAN jackpot and **never** eth `81` — meter did not reset. - **uartd:** `extract_local_x1_frame` requires gameplay-ACK **wire length 12** (same as `pog422_wire_len_primary` / `COMMAND_WIRE_LEN`); no first-EOP hunt. - **Python:** validate local wire-len ACK before eth-forward; on reject emit `inject_x1_ack_early_eth_reject` and **do not** latch `early_x1_eth_done`. Do not merge short X1_ACK into `post_de_rx`. - Driver still **2.2.2.10**. Rollback: `3.0.1.20` (x1-ack-eth-before-0d). ## 3.0.1.20 — 2026-08-01 — x1-ack-eth-before-0d Pos4 SpinBall progressive overnight (`2026-08-01_0318-UTC`, TID `37BC`, progwin **$3,126.80**): uartd peeled local `0x81` (`harvest=4` imm) and synth'd `0D` **before** Python could eth-forward the ack to Master. Hub saw Master `0x89` retries (`outcome_retry`) and **no `0x81` for `37BC`** — Slave showed Jackpot Won (progwin on `0x89`) while Master likely skipped progressive clear / pay accounting. - **uartd:** extract full local x1 SOH..EOP; emit `IPC_X1_ACK_FRAME` **before** synth `0D` inject; merge imm peel into luxury `POST_DE_RX` when post_de was skipped because x1 was already found. - **Python:** mid-inject `on_x1_ack_frame` eth-forwards ack immediately (during HS-idle wait / 0D TX); skip duplicate post-return early eth; merge x1 into `post_de` peer path. - Driver still **2.2.2.10**. Rollback: `3.0.1.19` (auto-close-skip-closed-seq). ## 3.0.1.19 — 2026-07-31 — auto-close-skip-closed-seq Pos5 SuperBall Keno MRT soak (`2026-07-31_1901-UTC_pos5-mrt-1611-edt-hs-timing-noroll`) on 3.0.1.18: peel/`synth_done` usually worked, but Master same-seq eth→UART `x9` barges after close still armed C auto-close → **1353/1392** false `status_0d_auto_close_miss harvest=0` while Python correctly logged `mrt_risk_arm_skip reason=seq_already_closed`. - **Python:** do not arm uartd auto-close when `play_sequence` is already in `_mrt_risk_close_done_seq`; emit `status_0d_auto_close_skip reason=seq_already_closed`. - Belt-and-suspenders: re-label a late miss the same way if close_done races. - uartd/inject unchanged (BOM tip string only). Driver still **2.2.2.10**. Rollback: `3.0.1.18` (auto-close-0d-wait-hs). ## 3.0.1.18 — 2026-07-31 — auto-close-0d-wait-hs Pos5 MRT ~09:46 EDT on 3.0.1.17 (`2026-07-31_1344-UTC_pos5-30117-mrt-after-restart`): immediate peel found local x1 (`harvest=4`) then synth `0D` failed with **EBUSY (16)** because HS was still high and `require_idle=0` means *reject if busy*, not force-TX. Python then `uartd_owns_close` → `late_0d_timeout`. - **uartd auto-close 0D:** `require_idle=1` + **150 ms** HS-idle wait — same collision model as other eth→UART traffic; **do not stomp** an active x1 window. - If 0D still fails after harvest, stash the built frame and **retry when HS falls** (`auto_close_0d_idle_retry_*`) within the pending window. - Python: label `harvest=4` as `immediate_post_x9`. Rollback: `3.0.1.17` (x9-immediate-rx-peel). ## 3.0.1.17 — 2026-07-31 — x9-immediate-rx-peel Pos1 MRT near midnight (`2026-07-31_1142-UTC_pos1-mrt-near-midnight-edt`): after a good local eth→UART x9 echo, uartd still returned `NO_X1` (`harvest=0`, `post_de_rx_len=0`) then re-injected Master's same-seq x9 retries at ~120 ms — holding the bus while the Slave's x1 (ms after hearing x9) was missed. Blind / infer `0D` is not the fix. - **uartd:** after armed local x9 DE-off, primary HS-agnostic RX peel: **20 ms first-activity gate** (any UART bytes), extend while RX continues (idle slice), complete on framed local x1, **45 ms hard cap**. `harvest=4` (`immediate_post_x9`). Zero activity in 20 ms → existing post_de / HS backup. - **Python:** defer same-seq eth→UART x9 while `await_x1` with `inject_deferred reason=await_x1_peel` (marks `retry_seen` so no-retry infer does not fire while Master keeps eth-retrying). - Driver unchanged (**2.2.2.10**). Rollback: `3.0.1.16` (hs-sample-tristate). ## 3.0.1.16 — 2026-07-30 — hs-sample-tristate Fix the `slv_hs_misalign` false-alarm storm introduced by 3.0.1.15. Archive `2026-07-30_2336-UTC_pos5-hs-de-hold-verify-30115` confirmed the DE fix held (1293 HS windows, max **26.9 ms**, zero > 100 ms) but logged **1442** `slv_hs_misalign` — exactly two per inject (`out_on` + `after_drain`), all `slv_in=false`. 3.0.1.15 replaced the DE-high `slv_in_phase_active()` samples with `gpio1_tr.last_active`, but nothing polled the trace between `gpio1_trace_init()` (pre-DE, seeds 0 on an idle line) and the snapshots, so the field was frozen low and the bridge scored every inject as a misalignment. Same false-alarm shape 2.2.1.14 fixed after 2.2.1.13, via a different route. - **inject:** new `gpio1_trace_poll_event()` — `GET_GPIO1_EVENT` only, which copies the driver's cached interrupt state and costs **no USB control transfer**, so it is safe while DE is high. Polled each drain-loop slice and at both snapshot points. - **Tri-state SLV-IN:** `POG422_SLV_IN_INACTIVE|ACTIVE|UNKNOWN` (0/1/2). DE-high samples report UNKNOWN when the GPIO1 event cache is off instead of masquerading as low. `slv_in_before` / `slv_in_after_off` stay boolean (both are sampled outside DE). - **Python:** `_slv_in_state()` maps 2 → `None`; bridge emits `slv_hs_misalign` only on a definite `False`. - No change to DE timing, bulk write, or capture. Driver stays **2.2.2.10**. Watch on the next soak: if the GPIO1 interrupt cache lags our own DE assert, misalign could still fire on a real-but-stale low. The Pi GPIO HS watcher (`hs_gpio_machine.jsonl`) is the better oracle and should own this alarm if so. Rollback: `3.0.1.15` (de-hold-hard-cap). ## 3.0.1.15 — 2026-07-30 — de-hold-hard-cap Kill the pos5 five-second SLV-OUT / HS stalls. Archive `2026-07-30_2211-UTC_pos5-hs-game-classes-no-mrt` proved both outliers were **our** MANUAL_DE: bulk/`tx_ns` finished in ~0.1–0.3 ms with `de_release_reason=temt`, yet `de_hold_us` stayed ~5.03–5.25 s until `gpio7_sl_out(0)` returned. Root cause: generic `xr_usb_serial_set_reg` / `usb_control_msg` uses a **5000 ms** timeout, and the DE-high path still called `capture_poll` / `GET_REG` (via `inject_poll_trace` and `slv_in_phase_active`) before release. A wedged control EP therefore held GPIO7 for a full five seconds — exactly the Python `INJECT_WAIT_S_DEFAULT` / USB ctrl ceiling — while the driver's between-chunk wall-clock never ran. - **Driver 2.2.2.10:** `pog422_set_gpio7_sl_out` uses a dedicated **100 ms** USB control timeout (not 5000). MANUAL_DE arms a delayed-work hard-cap at wire-air+20 ms (15–300 ms clamp); `SET_GPIO7(0)` and error paths disarm it. - **inject:** no `capture_poll` / GET_REG while DE is high; HS snapshot from the GPIO1 trace only; release GPIO7 immediately after TEMT/deadline. - **Python:** `INJECT_WAIT_S_DEFAULT` 5.0 → **0.75** (IPC hang ceiling, not DE). - Requires XR module **2.2.2.10** (`EG_COMPONENT_MIN_VERSION`). Rollback: `3.0.1.14` (clear-capture-keep-hs-gpio). ## 3.0.1.14 — 2026-07-30 — clear-capture-keep-hs-gpio Bench instrumentation fix. `CLEAR_CAPTURE=1` wiped **all** of `data/` at bridge startup, including a `hs_gpio_machine.jsonl` already opened by a foreground `hs_gpio_logger` in another terminal. Unlinking the path left the logger writing to an orphaned inode: it counted 1276 edges and printed a healthy summary, while `ls` and rclone saw nothing. Two pos5 HS-timing soaks were lost that way. - `clear_capture_files()` and `pog422_wipe_capture_dirs()` now preserve an `hs_gpio*` companion capture **only while a live process holds that inode open** (dev/inode scan of `/proc/*/fd`, not readlink text, since an unlinked fd reads back as `… (deleted)`). An unheld leftover is still deleted: keeping it would fold a previous run's edges into this session's archive, so the analyzer would correlate HS timing against the wrong `events.jsonl` — a silent wrong answer, worse than no data. Both outcomes are logged (`preserve (open by a live logger)` / `removed stale companion capture`). - `hs_gpio_logger.EdgeWriter` compares its fd inode against the path on every write/flush; if something unlinked or replaced the file it reopens and emits a `type=reopen` marker, so edges stay visible to sync. - `run_hs_gpio_logger.sh` sourced `${ROOT}/source_ethergold_env.sh`, which does not exist (env lives in `sync/`) — that is why the banner reported `pos=0`. Fixed, and the tap line now comes from `hs_gpio.env` (pos5 J19 = **GPIO3**) with a numeric guard, so a typo like `HS_GPIO_LINR=3` can no longer fall through to GPIO9. Covered by `tests/test_clear_capture_preserve.sh` (both wipe paths, held-open vs stale, inode identity, and the reopen marker) — run via `tests/run_all_unit.sh`. No wire, parser, or checksum change — capture-file lifecycle and diagnostics only. uartd/inject rebuild is required only to keep the BOM string equal to tip. Rollback: `3.0.1.13` (bank-jpot-grant-flood). ## 3.0.1.13 — 2026-07-30 — bank-jpot-grant-flood Extend `3.0.1.12` to the skill jackpot round trip. `3.0.1.12` keyed off `is_gameplay_response()`, which is `{0x29,0x39,0x49,0x59,0x69,0x79,0x89}` — it does **not** include the grants `0x5A`/`0x7A`, and those are exactly where Draw Poker royals and SpinJack bonus progressives carry `progwinval`. Those wins were still unicast to the winner's spoke, so other T340+ boards stayed silent. - `is_bank_progressive_jackpot_frame()` now accepts skill jpot grants (`is_skill_jpot_grant`) as well as outcome x9, so the Master's answer to a slave claim floods to every spoke. - `maybe_publish_bank_wlan()` had the same blind spot — skill grants never produced a clear-WLAN jackpot announce, so DyzineFrame fireworks also missed poker and SpinJack progressives on the star underlay. Grants now announce like an x9. - The slave claim itself (`0x52`/`0x72`) is deliberately not flooded: it carries no `progwinval` and is addressed to the Master. Still not covered: SpinBall Bonus `970D`. It is an 8-liner slot with no documented claim/grant pair (distinct from SpinJack `960F`, which is blackjack-based and does use `0x72`/`0x7A`), and `0x89` carries no `progwinval`. Note: the older bench `0x01` @ 17 B “poker skill reply” (v16aj) was a cropped packet aberration, not a real Master grant — do not treat it as a flood candidate. Rollback: `3.0.1.12` (bank-jpot-x9-flood). ## 3.0.1.12 — 2026-07-30 — bank-jpot-x9-flood Restore bank-wide T340+ bell on the wlan-star underlay. Since `2.2.2.71`, `EDGE_TID_UNICAST` pinned every gameplay x9 to the winner's spoke, so a progressive jackpot on one seat never reached the other T340+ boards and only the winner's cabinet could ring. On copper every slave saw that frame natively. - `is_bank_progressive_jackpot_frame()` in `pog422_core.py` — gameplay x9 whose `progwinval` (BE32 @ 13–16) is non-zero. Wager/ack/`0x0A` and zero-progwin outcomes are unaffected. - `_eth_only_ports_for_frame()` returns flood for those frames, so hub UART→ETH and ETH→ETH relay both reach every spoke. Ordinary gameplay stays unicast, so fan-out cost is unchanged except on an actual jackpot. - Defer queue no longer swallows a bell: `_abandon_transaction_deferred()` promotes bank jackpot frames to `_pending` (`source=bank_jpot_late`) instead of dropping them, and `queue_cap` eviction skips them. A late bell is acceptable; a missing one is not. - `transaction_defer_abandon` gains `kept_bank_jpot`. Slot progressives that carry no `progwinval` on `0x89` (SpinBall `970D`) are still not covered — that needs the symbol-map work in `SLOT_8LINER_EARLY_WIN.md`. Rollback: `3.0.1.11` (ohd-fireworks-http). ## 3.0.1.11 — 2026-07-29 — ohd-fireworks-http Expose the existing atomic `fireworks-request.json` event to Win11 OHD signs. - `ethergold-http` serves `GET /EtherGold/fireworks-request.json` from the same `ETHERGOLD_ROOT` used by the bridge and returns `{}` while no request exists (quiet normal polling, no repeated 404 access logs). - The endpoint is unauthenticated like `progressives.csv`; it contains only the existing display event payload and is polled by OHD on the trusted bank LAN. - No progressive detection, wire timing, UART, or inject behavior changed. Rollback: `3.0.1.10` (hs-gpio-ground-truth). ## 3.0.1.10 — 2026-07-29 — hs-gpio-ground-truth Bench instrument: Pi GPIO tap on `SLV_HS` so HS edge timing has a reference the USB paths cannot provide. Evidence-only — no wire-path or bridge behavior change. - `hs_gpio_logger.py` — drives the GPIO v2 character-device uAPI directly via `ioctl` (no libgpiod binding: Bookworm ships v1, Trixie v2). Kernel-stamped edges on `CLOCK_MONOTONIC`, the same base as uartd and the XR driver, so cross-source deltas are valid. Records `recv_ns` separately so delivery lag is measured rather than assumed, and emits `seq_gap` on event-buffer overflow. Matches the controller by **label** (`pinctrl-rp1` / `pinctrl-bcm2711`) because Pi 5 / CM5 renumbered the RP1 gpiochip between kernel releases. - `hs_gpio_analyze.py` — true HS width census (sub-3 ms buckets), real GET_REG poll lag, and how many HS-active windows each USB path misses outright. Excludes widths spanning a dropped-event gap. - `run_hs_gpio_logger.sh` — foreground launcher; writes `${OUTPUT_DIR}/hs_gpio_machine.jsonl` beside `events.jsonl` so `archive_session.ps1` picks it up. Touches no `/dev/ttyXRUSB*` and no uartd socket, so it runs concurrently with the bridge. - `HS_GPIO_TAP.md` — wiring (tap `SLV_HS` at R6/LED6, **not** `SLV-IN` — the TXS0108E one-shot stage is load-sensitive), J19 pin 7 = GPIO9 on the Seeed CM4 Dual GbE carrier, and the CM5/RP1 constraint that rules out pigpio. **`SLV_HS` is a 5 V signal** (HAT `VCC` = +5 V; only `VDD` is 3.3 V — the reason the TXS0108E exists). Pi GPIO is 3.3 V only and RP1 is not 5 V tolerant, so the tap requires a 74LVC1G17 buffer or a 1.8k/3.3k divider; a plain series resistor is not protection. Corrected before any hardware was built. Motivation, from `sessions/2026-07-29_1937-UTC_pos5-hs-timing-kernel-ring`: the XR Table-8 interrupt endpoint is hard-quantized to **8 ms** (every inter-arrival an exact multiple; HS widths are integer slot counts), and the GET_REG poll floors at ~2.75 ms with observed HS minimum 3.02 ms — the short-window population piles up at the instrument's own resolution limit. Both new scripts carry `--self-test` (uAPI struct/ioctl encodings; synthetic edge/width/miss logic) so they can be validated on the PC before deploying. Rollback: `3.0.1.9` (exar-only-fuser). ## 3.0.1.9 — 2026-07-29 — exar-only-fuser Stop EtherGold driver reload from killing DyzineFrame / PermDevMapper on the P3 FTDI cable. - `_stop_fixture_uart_users` in `sync/update.sh` only `fuser -k`s Exar nodes (`/dev/ttyXRUSB*` and `ttyUSB*` with udev `ID_VENDOR_ID=04e2`). - FTDI P3 (`0403:6001`), `/dev/ser_DT1`, and `/dev/pogcc` are never swept. Rollback: `3.0.1.8` (hs-timing-driver-floor). ## 3.0.1.8 — 2026-07-29 — hs-timing-driver-floor Make the timing-ring driver install automatic during the normal update. - Raise the XR driver floor from `2.2.2.8` to `2.2.2.9`; `update.sh` now sees 2.2.2.8 as out of range and installs/reloads 2.2.2.9 without `FORCE_DRIVER=1`. - Runtime timing remains opt-in through `POG422_TIMING_TRACE=1`; the 2.2.2.9 ring is dormant otherwise and does not change the wire path. - Deploy staging now validates the staged component floor against source instead of freezing a historical driver version in the build script. Rollback: `3.0.1.7` (hs-timing-env). ## 3.0.1.7 — 2026-07-29 — hs-timing-env Per-seat env-file switch for the 3.0.1.6 timing trace (pos5 only for now). - `sync/enable_hs_timing_env.sh` upserts `POG422_TIMING_TRACE` into `~/EtherGold/fixture-bridge.env`; refuses seats other than 5 / 51 without `--force`. - `source_ethergold_env.sh` exports `POG422_TIMING_TRACE*` (allowlist — otherwise uartd never saw the key from the env file). - uartd default trace path is now `${OUTPUT_DIR}/hs_timing_.jsonl` so it lands beside `events.jsonl` and `CHANNEL=both` daemons do not share one file. Rollback: `3.0.1.6` (hs-timing-trace). ## 3.0.1.6 — 2026-07-29 — hs-timing-trace Runtime-gated HS / RX / auto-close timing diagnostics (evidence only). - Driver `2.2.2.9`: always-compiled timing ring; `SET/GET/RESET/DRAIN_TIMING_TRACE` (off by default). - uartd: `POG422_TIMING_TRACE=1` → buffered JSONL (`data/hs_timing.jsonl`); no burst-on-fall change. - `hs_timing_analyze.py` + `HS_TIMING_DIAG.md` + fixture self-test. - Independent of `POG422_HW_DEBUG` (that mode still changes RX drain). Rollback: `3.0.1.5` (c-close-always). ## 3.0.1.5 — 2026-07-28 — c-close-always C owns local status `0D` after x1 — always (no Python uart_window fallback). - After inject `NO_X1`, keep pending x9 ≤200 ms; HS WINDOW/grace still hard-RT closes. - Bridge skips Python synth queue when `uartd_inject`; late `AUTO_CLOSE_DONE` bookkeeping. Rollback: `3.0.1.4` (auto-close-hs-x1). ## 3.0.1.4 — 2026-07-28 — auto-close-hs-x1 Pos5 Shamrock 7's MRT 14:56 EDT: copper `51` completed as `uart_window` while C auto-close returned `NO_X1` (echo=self-`59`, `post_de_len=0`); Python late `0D`. - Stash local-TID x1 from HS WINDOW flush while auto-close armed. - If echo+post_de miss, wait/drain HS burst or FIFO (≤40 ms) and peel (`harvest=3`). - C still injects synth `0D` before DONE on that path. Rollback: `3.0.1.3` (uartd-auto-reinit). ## 3.0.1.3 — 2026-07-28 — uartd-auto-reinit Field CGS after pos5 XRUSB/HS thrash: uartd died (`bad ipc magic 0x0` then zombie) while bridge kept running → forever `uartd not connected`. - Bridge **auto-respawns** pog422_uartd on IPC death / zombie (wait for tty, respawn binary, reconnect socket). Events: `uartd_auto_reinit`. - `uartd_common.sh` exports `POG422_UARTD_{PID,PORT,CHANNEL,LOG,BIN}` for adopt. Rollback: `3.0.1.2` (abort-de-hs-peel). ## 3.0.1.2 — 2026-07-27 — abort-de-hs-peel Pos5 17:37 MRT follow-up on uart-auto-close: - **IPC write deadline (100 ms):** never pin `ipc_mu` forever on a full client socket (DONE write blocked → GPIO7 abort ACK never sent → `aborted_de=0`). - **GPIO7 RELEASE retries** (3×) + best-effort ACK after DE drop. - **Auto-close HS peel:** keep incremental RX while SLV-HS already high after local x9 (was `burst_on_fall` → FIFO held until fall → `NO_X1`); arm fall capture after peel. - **Python:** abort RELEASE retries; longer orphan drain; `inject_abort_cooldown` before eth→UART requeue barge while HS still up. Rollback: `3.0.1.1` (uart-auto-close). ## 3.0.1.1 — 2026-07-27 — uart-auto-close Minor-feature BOM spin for under-the-hood hard-RT architecture: local close moves into uartd. After eth→UART x9, harvest peer x1 and inject synth `0D` in C before any Python IPC. `INJECT_DONE` precedes echo/forensics so `_tx_lock` cannot pin on Dropbox/logging (Class J / pos5 15:10 MRT: `inject_total_us≈19ms` vs `inject_ioctl_us≈886ms`). Python arms trailer + logs `status_0d_synth_done source=uartd_auto` after the fact. - BOM stamp tip → **3.0.1.1** (rebuild uartd/inject on Pi). - Same change briefly stamped `3.0.0.13` before this renumber. Rollback: `3.0.0.12` (current-game-catalog). ## 3.0.0.12 — 2026-07-27 — current-game-catalog - Expand `WIRE_GAME_TO_INDEX` from lab seed (7) to **13/16** titles from the 2026-06-19 wire-len census (`9400/01/03/04`, `9206`, `9107`, `9508`/`9507`, `9309`/`930A`, `970C`/`970D`, `960E`/`960F`). Still unmapped until capture: Wild Jokers, 8-Ball Poker, Lightning Keno. - BOM stamp tip → **3.0.0.12**. Rollback: `3.0.0.11` (current-game-ipc). ## 3.0.0.11 — 2026-07-27 — current-game-ipc - **Current-game IPC for DyzineFrame:** bridge maps local-TID M/S gameplay `game_id_hex` → DF `curr_index` (1..16) and publishes `~/EtherGold/current-game.json` only when the index changes (fireworks-style atomic JSON). - Env `ETHERGOLD_CURRENT_GAME_IPC`: `0` off (default), `readonly` compare/log only, `1` write-on-change. Events: `current_game_would_publish` / `current_game_published`. - Incomplete wire map (Touch6/Bingo/Keno/SuperPick/SpinJack/Triple7s/SpinBall); unmapped skip. - BOM stamp tip → **3.0.0.11**. Rollback: `3.0.0.10` (field-ui-drop-spe). ## 3.0.0.10 — 2026-07-27 — field-ui-drop-spe - Remove SPE from field config UI / profiles (LAN8670 multidrop is not galvanically isolated — does not solve mixed 3-phase common-mode). Stale `profile=spe` / `EDGE_UPLINK=spe` still rejected. - BOM stamp tip → **3.0.0.10**. Rollback: `3.0.0.9` (inject-abort-de-drop). ## 3.0.0.9 — 2026-07-27 — inject-abort-de-drop - **Kill pos5 02:51 MRT class:** orphan `INJECT_TX` after gameplay timeout left DE/HS high ~5 s; Python returned `None` with no abort; `GET_SLV_IN` blocked ~2 s. - **uartd:** inject on worker thread; main services `GPIO7_RELEASE` / `GET_SLV_IN`; concurrent inject → `EBUSY` DONE; early `IPC_INJECT_ECHO` when framed echo ready (arms Python `done_extend` before TEMT/DONE). - **Python:** timeout → `gpio7_release` + orphan IPC drain; gameplay requeue (`False`); `GET_SLV_IN` ≤50 ms; side-queue inject IPC; pre-inject HS from cache. - **Driver 2.2.2.8:** MANUAL_DE wall-clock — drop GPIO7 if bulk exceeds `timeout_ms+100ms` (floor `EG_COMPONENT_MIN_VERSION=2.2.2.8`). - BOM stamp tip → **3.0.0.9** (rebuild uartd/inject **and** XR module). Rollback: `3.0.0.8` (inject-log-after-tx). ## 3.0.0.8 — 2026-07-26 — inject-log-after-tx - **Inject critical path:** do not flush `inject_echo.jsonl` / `packets.raw` under `_tx_lock`. Queue forensics in memory; run peer harvest, synth `0D`, and pending injects first; write logs after close path. - Fixes pos4 20:55 EDT MRT class: ~1.2s Dropbox `flush` after a ~28ms UART inject delayed synth close while SLV markers looked like an orphan hold. - BOM stamp tip → **3.0.0.8** (rebuild uartd/inject on Pi for version match). Rollback: `3.0.0.7` (uart-xrusb-reopen). ## 3.0.0.7 — 2026-07-26 — uart-xrusb-reopen - **UART XRUSB reopen:** when Exar USB drops/renumbers (`ttyXRUSB1 (deleted)` while `0`+`2` remain), `pog422_uartd` detects dead fd, emits `uart_port_lost` / `uart_port_reopen`, re-resolves by channel (ois=lowest, machine=prefer `1` else highest), and reopens without killing Python/WG. - **`--port`** is first-open hint only; attach/reattach always use channel resolve. Launchers pass `--channel`. - **Queue hold:** eth→UART `_pending` defers while port down; unsuccessful transfers (ENODEV/EIO/timeout during loss) requeue — not terminal `inject_failed`. Events: `inject_requeued_uart_lost`. - BOM stamp tip → **3.0.0.7** (rebuild uartd/inject on Pi). Rollback: `3.0.0.6` (pipe-probe). ## 3.0.0.6 — 2026-07-26 — pipe-probe - **EDGE pipe probe:** bidirectional PING/PONG over WG+gretap (`EGPIPE1`, eth tag `P`) — **never** UART-injected to POG422. Retries until first success; records `pipe_ready_delay_s` / attempts / RTT; optional ongoing RTT samples. Events: `pipe_probe_start|tx|rx|retry|ready|rtt|miss`. Default on when `EDGE_MODE=1` (`PIPE_PROBE=0` to disable). - Addresses first-game aborts where spoke `eth_tx` ran but hub saw no uplink for ~15–20 s (see session eval 2026-07-26_1529). - BOM stamp tip → **3.0.0.6**. Rollback: `3.0.0.5` (field-config-ui). ## 3.0.0.5 — 2026-07-26 — field-config-ui - **Field configuration UI** on ethergold-http (`/EtherGold/config`): login (default SGVI), profile radios, save/validate/apply, `field-config.json` → `/etc/ethergold/{role,ethergold,edge}.env`. See `FIELD_CONFIG_MATRIX.md`. - **Identity:** free-text `asset_id` + bank; spoke unique ID = POG422 **TID**; no field `FIXTURE_POS` (derived internal). - **Always CHANNEL=both**; dual-channel **listen** (machine + OIS status hooks; OIS not forwarded to OHD progressives HTTP). - **OIS off VPN:** `ois_wifi_plane` HMAC flood on clear 6 GHz WiFi when `OIS_WIFI_PLANE=1`; inject with no RX TID filter. Machine stays on WG+gretap. - **Atlas star + DyzineFrame:** `set_ethergold_role.sh edge --display-on` separates transport from display. - BOM stamp tip → **3.0.0.5**. Rollback: `3.0.0.4` (master-uart-x9-no-synth). ## 3.0.0.4 — 2026-07-25 — master-uart-x9-no-synth - **Master vs Slave synth 0D:** track per-round x9 origin (`uart` vs `eth`). UART-origin local x9 (Master native gameplay response) ⇒ **never** synth 0D — POG422 closes with native UART `0x0D`. Eth-injected local x9 (Slave) still arms synth after local x1. Skip reason: `master_uart_x9_native_0d`. - BOM stamp tip → **3.0.0.4**. Rollback: `3.0.0.3` (hub-peer-inject). ## 3.0.0.3 — 2026-07-25 — hub-peer-inject - **Hub artificial freeze:** with `EDGE_TOPOLOGY=hub`, one local close (`await_local_0d` / up to **2 s** `await_local_x1`) was blocking **all** other TIDs' eth→UART gameplay — every other spoke waited in `transaction_deferred`. Soft **0x0A** still defers (latest-wins). Other spokes' gameplay now proceeds; same-TID close path + P0 synth 0D unchanged. Copper single-Master (no hub topology) keeps the old serialize-on-close behavior. - Soak evidence: spoke req→reply p50≈125–150 ms, p90≈1 s+, unmatched CGS; hub `transaction_defer await_local_0d` and inject-hold were the pipe stalls — not gretap fan-out (p90≈0.12 ms). - BOM stamp tip → **3.0.0.3**. Rollback: `3.0.0.2` (spoke-tunnel-fix). ## 3.0.0.2 — 2026-07-25 — spoke-tunnel-fix - **Spoke launch crash:** `tunnel_up` mis-read gretap endpoints (`ip -d` prints `remote…local…`; checker wanted `local…remote…`), treated a good `eg-sNN` as mismatch, then fell through to legacy `gretapN` and died on `gretap0@NONE` / `File exists`. Spokes now use `ensure_spoke_gretap` only. - **pos4 leftover `bench.env`:** `POG_EDGE_ROLE=pos4` overrode `EDGE_ROLE=spoke`. Canonical `EDGE_TOPOLOGY=hub|spoke` from `/etc/ethergold/edge.env` now wins; provision retires single-hop `bench.env`. - **Stale live WG:** when `wg0` already up, hub/spoke syncs from `wg0.conf` (`wg syncconf` / bounce) instead of only patching Endpoint on the old peer. - BOM stamp tip → **3.0.0.2**. Rollback: `3.0.0.1` (hub-spoke). ## 3.0.0.1 — 2026-07-25 — hub-spoke - **Major BOM cut:** wlan-star hub/spoke underlay is the product line for all-wireless banks (no inter-seat eth). Prior `2.2.2.71`–`2.2.2.75` work is the development path; field/tip BOM is now **`3.0.0.1`**. - **Topology:** `EDGE_MODE=1` + `EDGE_TOPOLOGY=hub|spoke`; Master = hub, Slaves = spokes with direct WG tunnels; TID-unicast gameplay + clear-WLAN meters/jackpots. - **One edge.env:** `/etc/ethergold/edge.env` canonical; `/etc/fixture/edge.env` is a symlink on provisioned seats. One command per seat: `cd ~/EtherGold/dev && bash run_ethergold.sh`. - **Provision:** `provision_hub_spoke.sh` + `sync/provision_wlan_star.ps1` pre-seed WG identity / `peers.json` so a zero-peer hub can bind `eg-sNN` on first start. - XR driver still OK in **`[2.2.2.7, 3.0.0.1]`** (range check). uartd/inject must equal tip. - BOM stamp tip → **3.0.0.1**. Rollback: `2.2.2.75` (edge-env-canonical) for emergency tip only. ## 2.2.2.75 — 2026-07-25 — edge-env-canonical - **Single edge.env — no fork:** `/etc/ethergold/edge.env` is the one canonical config. `/etc/fixture/edge.env` becomes a **symlink** to it on provisioned seats (`.74`'s dual-write mirror retired — two real files can diverge). `run_ethergold.sh` / `pog_edge_env.sh` back to single-source: canonical if present, else legacy plain-file fallback for unmigrated boxes. - **provision_wlan_star.ps1** writes only the canonical file, parks any real legacy file as `edge.env.retired.`, and installs the symlink. - **apply_chain_machine_uartd.sh** CRLF strip uses `sed --follow-symlinks` (plain `sed -i` would replace the legacy symlink with a diverging copy). - BOM stamp tip → **2.2.2.75**. Rollback: `2.2.2.74` (edge-env-overlay). ## 2.2.2.74 — 2026-07-25 — edge-env-overlay - **One-command star:** `run_ethergold.sh` / `pog_edge_env.sh` source `/etc/ethergold/edge.env` **then** `/etc/fixture/edge.env` (fixture wins). Stale `EDGE_MODE=0` in the ethergold path no longer silently drops hub/spoke config written to fixture. Banner shows `topo=` / `uplink=`. - **provision_wlan_star.ps1** writes **both** edge.env paths so the next provision lands where `run_ethergold` looks first. - Operator path remains: `cd ~/EtherGold/dev && bash run_ethergold.sh` (hub first). No `FIXTURE_POS=` / `run_bridge_edge_bench.sh` prefix needed — `fixture-bridge.env` + edge.env drive it. - BOM stamp tip → **2.2.2.74**. Rollback: `2.2.2.73` (hub-spoke-provision). ## 2.2.2.73 — 2026-07-25 — hub-spoke-provision - **Bootstrap fix (provisioning, not parser):** a hub with zero enrolled spokes had no gretap to bind — `tunnel_up.sh` fell back to a placeholder `eg-s02` that does not exist, and the enrollment listener only starts *inside* the bridge. New `pog-edge/provision_hub_spoke.sh` seeds hub `[Interface]` conf, per-spoke `wg0.conf` (hub pub + endpoint), and `peers.json` / `allowlist.txt`, so `restore_hub_spoke_gretaps` has real `eg-sNN` legs on first start. - **PC orchestrator:** `sync/provision_wlan_star.ps1` patches `edge.env` per seat (`EDGE_TOPOLOGY`, `EDGE_SPOKE_ID`, `EDGE_GAME_IF=none`), collects keys, seeds the hub. Config + keys only — operator still starts bridges in the foreground. - Spoke IDs default to `10 + seat` so the overlay last octet reads as the seat (pos1 → `10.99.0.11` / `eg-s11`). - Live enrollment path unchanged: it now covers re-enroll and virgin seats rather than first-star bootstrap. - BOM stamp tip → **2.2.2.73**. Rollback: `2.2.2.72` (enroll-crypto-stealth). ## 2.2.2.72 — 2026-07-25 — enroll-crypto-stealth - **Enrollment replies:** bad HMAC / garbage stay **silent** (no "service exists" clue for probes). After a valid fleet HMAC, the hub always replies — accept **or** signed diagnostic deny (`bank_mismatch`, `window_closed`, `not_allowlisted`, `missing_pub`, `apply_failed`) so commissioning seats know what to fix. - BOM stamp tip → **2.2.2.72**. Rollback: `2.2.2.71` (hub-spoke-underlay). ## 2.2.2.71 — 2026-07-25 — hub-spoke-underlay - **Underlay layering:** `EDGE_TOPOLOGY=hub|spoke` extends `EDGE_MODE=1` without breaking `EDGE_MODE=0` eth↔eth or single-hop eth↔wlan (`peer-a`/`peer-b`). SPE (`EDGE_UPLINK=spe`) reserved until >8-node SPE confirmed. See `pog-edge/EDGE_HUB.md`. - **TID-unicast + fan-out metrics:** `DirectionalEthRouter` learns TID→port; gameplay unicasts when known; `eth_fanout` logs `wall_us` / mode for N=2/4 study. - **Clear-WLAN bank path:** signed `0x0A` meters + `progwinval>0` jackpot announces on non-VPN uplink (`pog_edge_bank_wlan.py`); discovery piggybacks on meter envelope. - **Identity:** read `MACHINE_ROLE`/`BANK_NUMBER`/`TERM_NUMBER` from DyzineFrame `config.ini` with process latch (`pog_edge_identity.py`) — no second `/dev/ttyUSB0`. - **Enrollment:** spoke keygen + hub UDP listener; live `wg set` + deterministic `eg-sNN` gretaps; persist peers; thumb-drive fallback (`pog_edge_enrollment.py`). - Multi-gretap: stop deleting mismatched sibling gretaps; hub `ETH_PORTS` from live spoke set; `FIXTURE_POS` gate `1..99` on EDGE bench launcher. - BOM stamp tip → **2.2.2.71**. Rollback: `2.2.2.70` (echo-armed-wait). ## 2.2.2.70 — 2026-07-25 — echo-armed-wait - **Echo-armed inject wait:** the gameplay 0.55 s budget still fails fast when no bus evidence arrives (18:33 anti-pin), but once echo / post_DE / forensics prove the TX left the driver, `inject()` extends **once** (`INJECT_DONE_EXTEND_S` = 1.0 s) for the late `IPC_INJECT_DONE`. Fixes pos5 2026-07-25 00:09 MRT: `79` delivered (`echo_len=42`), DONE ~84 ms past budget → false `uartd_inject_timeout` → Master retry → infer close. - **Socket honors the budget:** per-read recv timeout tracks remaining budget (idle bus no longer rides the fixed 2 s recv timeout past 0.55 s and surfaces as a bare `timed out`). Timeout detail now reports elapsed, budget, and `done_extend` state. - BOM stamp tip → **2.2.2.70**. Rollback: `2.2.2.69` (peel-inject-budget). ## 2.2.2.69 — 2026-07-24 — peel-inject-budget - **Bidirectional peel:** `peer_uart_bytes_after_inject` keeps peer bytes *before* and *after* self-echo; `inject_echo_bytes_match` adds `embedded` (pos5 18:33 leading `61|x9` was discarded → infer). - **Inject budget:** gameplay eth→UART / synth `0D` IPC wait **0.55 s** (`INJECT_WAIT_S_GAMEPLAY`); timeout fails without arming await/infer (stops multi-second `_tx_lock` pin). Slow-inject diag includes drain/TEMT proxies. - **Close policy:** `late_0d` / `already_sent` keep pending when P0 `0D` is queued or mid-flight; `_round_close_phase` log label. - BOM stamp tip → **2.2.2.69**. Rollback: `2.2.2.68` (x1-always-0d). ## 2.2.2.68 — 2026-07-23 — x1-always-0d - **Always synth `0D` after a real local `x1`:** remove `status_0d_synth_skip` `no_open_await`. Infer/premature `0D` is fine; skipping close when copper/`uart_window` `61` arrives after seq already closed was the 18:07 EDT MRT hole. `_mark_local_x1_pending` still releases `synth_sent` so a second close can TX. - BOM stamp tip → **2.2.2.68**. Rollback: `2.2.2.67` (infer-0d-25ms). ## 2.2.2.67 — 2026-07-23 — infer-0d-25ms - **Harvest-miss `0D` before slave MRT:** `MRT_X1_INFER_NO_X9_RETRY_MS` and `MRT_X1_INFER_AFTER_X9_RETRY_MS` **125/100 → 25 ms**. Slave arms its `0D` watchdog when it sends `61` (after we drop DE); waiting 125 ms to infer+synth was itself an MRT (pos5 16:57 EDT / w51). Real harvested x1 still closes immediately via `.66` early queue. - BOM stamp tip → **2.2.2.67**. Rollback: `2.2.2.66` (early-0d-queue). ## 2.2.2.66 — 2026-07-23 — early-0d-queue - **0D must arm on x1, not on inject DONE:** while eth→UART x9 waits for `INJECT_DONE`, `INJECT_ECHO` / `POST_DE_RX` callbacks harvest local x1 and `appendleft` synth `0D` onto `_pending` (even under `_tx_lock`). - `_schedule_synth_attempt` always builds/dispatches (dispatch queues under lock); no longer only parks an rk until lock release. - uartd: skip post-DE poll for non-x9 (stops 25+15 ms stall on every `0D`); skip post-DE when DE-high echo already has a peer EOP after self. - Synth inject: `idle_wait_ms=0` (HS busy → requeue); skip post-HS settle / OIS post-x9 grace for P0 close; process pending 0D at end of `send_frame`. - BOM stamp tip → **2.2.2.66**. Rollback: `2.2.2.65` (edge-bench-entry-guard). ## 2.2.2.63 — 2026-07-22 — deploy-dev-sync-only - `To Deploy/EtherGold` is only `dev/` + `sync/` + `install_field.sh`. Dropped legacy `bridge/` and `Daemon/` duplicates. - Field `~/EtherGold/prod` is built on the Pi via `build_prod.sh`. - BOM stamp tip → **2.2.2.63**. Rollback: `2.2.2.62` (no-windows-in-deploy). ## 2.2.2.62 — 2026-07-22 — no-windows-in-deploy - `To Deploy/EtherGold` excludes Windows-only `*.ps1` / `*.bat` / `*.cmd` (PC sync helpers stay in workshop `EtherGold/sync` only). - rclone filter also blocks those extensions on Pi pull. - BOM stamp tip → **2.2.2.62**. Rollback: `2.2.2.61` (pull-from-to-deploy). ## 2.2.2.61 — 2026-07-22 — pull-from-to-deploy - Bench pull root is Dropbox `To Deploy/EtherGold` (`dev/` + `sync/`). - `build_to_deploy.ps1` stages those folders; `update.sh` pull migrates workshop `EtherGold` → To Deploy in `rclone.env`. - BOM stamp tip → **2.2.2.61**. Rollback: `2.2.2.60` (update-run-ethergold). ## 2.2.2.60 — 2026-07-22 — update-run-ethergold - `update.sh` prints one start line: `bash run_ethergold.sh` (role/EDGE from env). Optional `UPDATE_START=1` execs it after rebuild. - BOM stamp tip → **2.2.2.60**. Rollback: `2.2.2.59` (cli-from-env). ## 2.2.2.59 — 2026-07-22 — cli-from-env - Operator start is only `bash run_ethergold.sh` — role, EDGE, POS, CHANNEL, and prod/bench come from env files; EDGE auto-sudo; listen dispatched from `ETHERGOLD_ROLE=listen`. - `source_ethergold_env.sh` exports `ETHERGOLD_ROLE` / `ETHERGOLD_PROFILE`. - BOM stamp tip → **2.2.2.59**. Rollback: `2.2.2.58` (env-pos-autoload). ## 2.2.2.58 — 2026-07-22 — env-pos-autoload - Launchers load `FIXTURE_POS` / `CHANNEL` from `fixture-bridge.env` (and `/etc/ethergold`) — no need to prefix `FIXTURE_POS=N` on the command line. - Touches: `run_ethergold.sh`, `run_bridge_prod.sh`, `run_listen_prod.sh`. - BOM stamp tip → **2.2.2.58**. Rollback: `2.2.2.57` (unified-bridge-entry). ## 2.2.2.57 — 2026-07-22 — unified-bridge-entry - Single operator entry: `run_ethergold.sh` dispatches copper vs EDGE from `EDGE_MODE` + `ETHERGOLD_PROFILE` (prod|bench). - `run_bridge_prod.sh` / `run_bridge_jun11_good.sh` redirect into `run_ethergold.sh` when `EDGE_MODE=1` (unless `ETHERGOLD_ENTRY` already set), so copper launchers cannot drop gretap on EDGE seats. - `ethergold_supervisor.py` starts `run_ethergold.sh` instead of branching scripts by hand. - BOM stamp tip → **2.2.2.57**. Rollback: `2.2.2.56` (shim-private-reexport). ## 2.2.2.56 — 2026-07-20 — shim-private-reexport - `pog422_fixture_core` shim re-exports private `_` names (`import *` skips them). Fixes `--echo-events` flood of bare `{"ev":"error"}` from `cannot import name '_clear_post_x9_grace_state'`. - `pog422_uart_capture` imports from `pog422_core` directly. - Echo brief includes `where` / `message` / `errno` so errors are diagnosable. - BOM stamp tip → **2.2.2.56**. Rollback: `2.2.2.55` (prod-role-ethergold-rename). ## 2.2.2.55 — 2026-07-20 — prod-role-ethergold-rename - Field roles: `ETHERGOLD_ROLE=topbox|listen|edge` via `set_ethergold_role.sh` / `commission_prod_boot.sh` (default **topbox**). - Paths: `/etc/ethergold`, `ethergold.service` / `ethergold-collector.service`, `/var/lib/ethergold` (legacy `/etc/fixture` + `fixture-prod` migrated once). - Module rename (compat shims kept): `pog422_bridge` / `pog422_core` / `pog422_listen`, `hat_bridge` / `hat_listen`, `ethergold_*_supervisor`. - BOM stamp tip → **2.2.2.55**. Rollback: `2.2.2.54` (superpick-9508-replicate). ## 2.2.2.54 — 2026-07-20 — superpick-9508-replicate - SuperPick synth: bus id **9508** (floor); capture-replicate wager (TID+game+seq+CRC); stop random pick rewrite that yielded Master-silent `0x60`. - BOM stamp tip → **2.2.2.54**. Rollback: `2.2.2.53` (lab-seat-close-on-x1). ## 2.2.2.53 — 2026-07-20 — lab-seat-close-on-x1 - Lab seat SM: round closes when forged **x1** is sent (no skill/jpot, no 0D synth). Inbound Master 0D still accepted by normal bridge path. - BOM stamp tip → **2.2.2.53**. Rollback: `2.2.2.52` (lab-seat-sm-eth-rx). ## 2.2.2.52 — 2026-07-20 — lab-seat-sm-eth-rx - Lab seat SM: also arm x1 ack on **eth_rx** of Master x9 (not only after eth→UART inject) — mid-chain synth hops may not complete copper inject. - BOM stamp tip → **2.2.2.52**. Rollback: `2.2.2.51` (lab-gameplay-seat-sm). ## 2.2.2.51 — 2026-07-20 — lab-gameplay-seat-sm - Lab gameplay synth is a **full seat state machine** for all catalog games: forge wager (x0) → on Master outcome (x9) forge ack (x1) → bridge 0D close. - Applies to touch6 / bingo / keno / poker / superpick / spinjack / triple7s (`0x2x`…`0x8x` families). - Dev Lite: `lab_gameplay_synth_ack`. - BOM stamp tip → **2.2.2.51**. Rollback: `2.2.2.50` (lab-synth-env-preserve). ## 2.2.2.50 — 2026-07-20 — lab-synth-env-preserve - `run_bridge_jun11_good.sh`: preserve CLI `LAB_GAMEPLAY_*` across `source_fixture_env` (fixture-bridge.env was forcing `LAB_GAMEPLAY_SYNTH=0`). - BOM stamp tip → **2.2.2.50**. Rollback: `2.2.2.49` (touch6-autoplay-golden). ## 2.2.2.49 — 2026-07-20 — touch6-autoplay-golden - Refresh Touch6 golden from **pos2 full-Dev autoplay** (`packets.raw`, TID `38D7`, 1 card, `byte8=0x04`) — replicate still TID+seq+CRC only. - BOM stamp tip → **2.2.2.49**. Rollback: `2.2.2.48` (touch6-golden-replicate). ## 2.2.2.48 — 2026-07-20 — touch6-golden-replicate - Touch6 lab synth default: **capture-replicate** copper golden (`lab_gameplay_templates/touch6_wager.hex` from pos2 `2026-07-03_2212-UTC_touch6-jpot-3game-success`) — rewrite **TID + seq + CRC only**. - Allow any 4-hex `LAB_GAMEPLAY_TID` (pilot alts: `E001`, `40CC`). - Extract helper: `lab_gameplay_extract_touch6.py` from full-Dev `packets.raw`. - BOM stamp tip → **2.2.2.48**. Rollback: `2.2.2.47` (lab-gameplay-msg-fix). ## 2.2.2.47 — 2026-07-20 — lab-gameplay-msg-fix - Fix `lab_gameplay_synth_wager` emit: duplicate `msg=` kwarg crashed the synth thread after the first UART→ETH forge (no further rounds). - BOM stamp tip → **2.2.2.47**. Rollback: `2.2.2.46` (lab-gameplay-synth). ## 2.2.2.46 — 2026-07-20 — lab-gameplay-synth - **Lab synthetic T340+ gameplay inject** (bus-load soak): `lab_gameplay_synth.py` forges slave wager frames as UART→ETH origin on a reserved lab TID (`E000`–`E0FF`). Rotating ~5 min game sessions; within session bet tiers min→medium→high from `Game_Bet_Tiers.xlsx` (25¢ credits). Catalog: touch6, superpick, bingo, poker (default) + keno/spinjack/triple7s. Master still serves real outcomes; local `0D` synth unchanged. Orthogonal to Master JPOT egress hijacks. - Env: `LAB_GAMEPLAY_SYNTH`, `LAB_GAMEPLAY_TID`, `LAB_GAMEPLAY_GAMES`, `LAB_GAMEPLAY_SESSION_S`, etc. Prod hard-refuse unless `LAB_GAMEPLAY_SYNTH_ALLOW_PROD=1`. - Dev Lite events: `lab_gameplay_session_start|end`, `synth_wager`, `synth_skip`. - Docs: `LAB_GAMEPLAY_SYNTH.md`. - BOM stamp: `pog422_ipc.h` + `pog422_uart_client.py` → **2.2.2.46** (tip align; no uartd behavior change). Rollback: `2.2.2.45` (dev-light-outcome-eth). ## 2.2.2.45 — 2026-07-20 — dev-light-outcome-eth - **Dev-light soak crumbs** (WiFi-edge / MRT eval gaps): keep Lite overnight, add thin gameplay wire visibility without full Dev / `packets.raw`. - Allow `eth_tx` + `eth_to_uart_done` in Dev Lite for wager / outcome / x1 / `0D` (still suppress `0A` flood). - `inject_peer_early_eth` now includes `peer_msgs`. - New `peer_gameplay_eth_fwd`: each gameplay response/ACK/`0D` in early-forwarded inject peer (first Master outcome TX path after `50` inject). - UART→ETH gameplay `eth_tx` adds `outcome_retry` + `retry_gap_ms` when same TID+seq+msg reappears within 2.5 s (Master `59` resent). - Python-only — no uartd rebuild required for behavior; BOM stamp tip align. - BOM stamp: `pog422_ipc.h` + `pog422_uart_client.py` → **2.2.2.45**. Rollback: `2.2.2.44` (eth-thread-ipc-timeout). ## 2.2.2.44 — 2026-07-19 — eth-thread-ipc-timeout - **Pos4 lockup fix:** eth→UART threads no longer die on exceptions; log and continue. - `eth_rx_queued` uses **cached** SLV-IN (no blocking IPC) for the idle diagnostic field. - `UartDaemonClient.slv_in_active`: on IPC stall/timeout return cache instead of raising `TimeoutError` (pos4 `.43`: dual eth thread killed → Game 4 `80` retry storm). - Keep `.43` infer @100 ms + `.42` orphan gate + `.41` harvest. - BOM stamp: `pog422_ipc.h` + `pog422_uart_client.py` → **2.2.2.44**. Rollback: `2.2.2.43` (infer-retry-100ms). ## 2.2.2.43 — 2026-07-19 — infer-retry-100ms - **Infer after same-seq x9 retry:** drop miss-backstop from **250 ms → 100 ms** (near empirical slave MRT window). Harvested local x1 still schedules synth `0D` immediately (no infer timer). - Keep `.42` orphan `solo_x1` gate + `.41` harvest stack. - BOM stamp: `pog422_ipc.h` + `pog422_uart_client.py` → **2.2.2.43**. Rollback: `2.2.2.42` (await-gate-infer-retry). ## 2.2.2.42 — 2026-07-19 — await-gate-infer-retry - **Orphan solo_x1 gate (pos1 20:24 MRT):** UART `solo_x1` / glued harvest only marks pending when `await_x1`, `await_0d`, or late-miss `logged_x1` is open. Blocks second synth `0D` after `seq_already_closed` (`status_0d_synth_skip reason=no_open_await`). - **Infer after same-seq x9 retry (pos1 overnight T2):** if `x9_retry_seen` still has no local x1 by **250 ms**, allow infer (`after_x9_retry=true`) instead of waiting for the 3 s `await_local_x1_timeout`. 125 ms no-retry infer unchanged. - Keep `.41` harvest stack (Class A/B/D). Infer still present as backstop. - BOM stamp: `pog422_ipc.h` + `pog422_uart_client.py` → **2.2.2.42**. Rollback: `2.2.2.41` (x1-harvest-ipc-queue). ## 2.2.2.41 — 2026-07-18 — x1-harvest-ipc-queue - **Class A:** `UartDaemonClient` side-queues `WINDOW`/`HS_*`/`DISCARD` while `inject()` (and other RPC waiters) hold the socket — drained via `poll()`/`drain_events()` after DONE so pre-inject flush x1 is not silent-dropped. - **Class B:** `pog422_uartd` re-arms 25 ms post-x9 grace after `flush_post_de_rx` so grace no longer expires during the post-DE poll window. - **Class D:** mismatch peel finds embedded self-echo substring; peer local-x1 prefers TID-checked packets; `mrt_x1_after_x9` logs `msg`+`trigger`; new `x1_harvest_path` event (`uart_window|post_de_peel|grace|infer`). - Instrumentation: stderr `dttty_sink=…` + `post_de_vs_hs_steal`; DEV_LIGHT adds harvest/grace/peer events. Infer **kept**. - late_0d / await_0d abandon clears `_status_0d_synth_sent`. - BOM stamp: `pog422_ipc.h` + `pog422_uart_client.py` → **2.2.2.41**. Rollback: `2.2.2.40` (narrow-tx-lock-fast-0d). ## 2.2.2.40 — 2026-07-18 — narrow-tx-lock-fast-0d - **Keep** `.39` x1-infer (not removed until harvest root cause is gone). - Fix pos2 08:04 EDT late-synth MRT class: `_tx_lock` now covers **uartd.inject only**; peer parse / x1 mark / remember / packet log / synth run **after** lock release so closing `0D` is not stalled behind inject epilogue I/O. - Log `status_0d_synth_deferred_tx_lock` when synth is queued because lock is held. - Log `mrt_x9_inject_slow` when inject_total_us ≥ 100 ms (forensics). - BOM stamp: `pog422_ipc.h` + `pog422_uart_client.py` → **2.2.2.40**. Rollback: `2.2.2.39` (x1-infer-no-x9-retry). ## 2.2.2.39 — 2026-07-17 — x1-infer-no-x9-retry - After armed local x9: if Master does **not** re-send same-seq x9 within **125 ms**, infer that local x1 was sent (we missed harvest) and synth close — `mrt_x1_inferred_no_x9_retry` / `status_0d_synth_trigger trigger=no_x9_retry`. - Same-seq Master x9 while awaiting x1 (`await_x1_same_seq`) marks retry seen → **no** infer; keep waiting for real x1 or 3 s `await_local_x1_timeout`. - Fixes pos4 soak class: copper had `81`, bridge mismatch-echo miss, Master retry only at ~2 s (after slave MRT window). - BOM stamp: `pog422_ipc.h` + `pog422_uart_client.py` → **2.2.2.39**. Rollback: `2.2.2.38` (tx-gate-cleanup). ## 2.2.2.38 — 2026-07-17 — tx-gate-cleanup - Real transaction-gate cleanup after overnight soak (pos1 late close + pos4 x9/no-x1 with defer queues >1000): - Cap `_transaction_deferred` at **64**; drop oldest on overflow (`transaction_defer_abandon` / `queue_cap`). - On `x9_no_local_x1` (3 s): emit `await_local_x1_timeout`, clear await_x1, close local trunk, **drop** deferred foreign (no flush storm). Late machine x1 can still arm await_0d via `logged_x1`. - Stale open trunk rounds abandon after **8 s** (`trunk_round_stale_abandon`) so `game_transaction_gate` / 0A cannot stick overnight. - Keep `.36/.37` defer/skip log throttling. - BOM stamp: `pog422_ipc.h` + `pog422_uart_client.py` → **2.2.2.38**. Rollback: `2.2.2.37` (relay-skip-log-throttle). ## 2.2.2.37 — 2026-07-16 — relay-skip-log-throttle - Rate-limit `status_0d_relay_synth_skip` (`local_same_game_active`) to ≤1/s. Guard behavior unchanged — expected under multi-seat same-game soak (foreign 50/59/51/80/89/81 for the same game while local TID already has state). - BOM stamp: `pog422_ipc.h` + `pog422_uart_client.py` → **2.2.2.37**. Rollback: `2.2.2.36` (tx-defer-log-throttle). ## 2.2.2.36 — 2026-07-16 — tx-defer-log-throttle - Rate-limit `transaction_defer` console/dev-light logs to ≤1/s per reason (`game_transaction_gate`, `await_local_0d`, `0a_latest_wins`). Defer queue behavior unchanged — only log spam under multi-seat soak. - BOM stamp: `pog422_ipc.h` + `pog422_uart_client.py` → **2.2.2.36**. Rollback: `2.2.2.35` (x1-synth-nonblocking). ## 2.2.2.35 — 2026-07-16 — x1-synth-nonblocking - Fix residual pos5 MRT on `.34`: UART local-x1 synth no longer calls `send_frame` while eth→UART holds `_tx_lock` (that blocked `capture_tick` ~900 ms and starved late-0D / retries). Queue P0 `status_0d_synth` / `_synth_retry_rks` instead; drain after lock release and every capture tick. - Every `_mrt_risk_tick`: re-arm `x1_pending` for any local `await_0d` and retry synth (heals await_passthrough until x9 hint arrives, and note-without-schedule races). - `superseded_seq`: if local `last_x9` already matches `latest_seq`, refresh hint and continue instead of abandoning. - Keep `x1_pending` until synth inject succeeds; in-flight queued synth is not treated as `already_sent` abandon. - BOM stamp: `pog422_ipc.h` + `pog422_uart_client.py` → **2.2.2.35**. Rollback: `2.2.2.34` (local-x1-synth-guard). ## 2.2.2.34 — 2026-07-16 — local-x1-synth-guard - Force local x1 close handling to mark `x1_pending` and schedule synth `0x0D` even when the x1 is seen in `peer_x1_in_echo`; the actual inject is deferred until the tx lock releases so the close cannot silently wait for a later Master retry. - Guard same-game foreign relay synth: when the local seat already has state for that game (e.g. pos1 and pos5 both playing Jacks or Better), foreign TID x9/x1 no longer rekeys into the local `(tid, game)` close state. - Dev-light now logs `status_0d_synth_trigger`, `local_round_superseded`, `status_0d_relay_synth_skip`, and `status_0d_synth_skip` reasons `superseded_seq` / `await_passthrough`. - BOM stamp: `pog422_ipc.h` + `pog422_uart_client.py` → **2.2.2.34**. Rollback: `2.2.2.33` (inject-tip-match). ## 2.2.2.33 — 2026-07-16 — inject-tip-match - Tighten policy after pos5 twice stayed on uartd **2.2.2.29** while Python was **2.2.2.32**: `update.sh` rebuilds inject/uartd when binary ≠ `EG_VERSION`; `pog422_ipc.h` must match tip (FATAL if stale); `session_start` `uartd_version_ok` / `inject_version_ok` require exact tip. - BOM stamp: `pog422_ipc.h` + `pog422_uart_client.py` → **2.2.2.33**. - XR driver still uses the component range floor. Rollback: `2.2.2.32` (await-0d-local-only). ## 2.2.2.32 — 2026-07-16 — await-0d-local-only - Fix CGS Master-not-responding from 2.2.2.31: foreign solo `x1` armed sticky `await_0d`, locking absolute P0 defer of SCR/STOM. Arm close-wait only for local TID; abandon on synth-fail / late-0D timeout; ignore non-local keys in `_local_close_pending`. - New event: `mrt_risk_await_0d_abandon`. - Python-only — no uartd rebuild. Rollback: `2.2.2.31` (await-0d-priority-dev-light). ## 2.2.2.31 — 2026-07-16 — await-0d-priority-dev-light - Absolute P0 priority for local synth `0x0D`: while `await_0d` / `status_0d_synth` queued, defer **all** other eth→UART (foreign TIDs, STOM/SCR) so close is never stuck behind relay inject. - Dev-light logging (`FIXTURE_DEV_LIGHT` / `--dev-light`): fault + MRT/synth `events.jsonl`; no `packets.raw`. Soak default on `run_bridge_jun11_good.sh`. Skip debug log rotate on inject hot path. - Python-only — no uartd rebuild. Rollback: `2.2.2.30` (early-x1-eth-forward). ## 2.2.2.25 — 2026-07-14 — empty-echo-reinject-no-arm - Local eth→UART `x9` with empty inject echo + empty post-DE: one immediate re-inject (`mrt_x9_empty_echo_reinject`); do **not** arm the 400 ms `x9_no_local_x1` miss timer or post-x9 grace/soft-hold until bus RX proves delivery (`mrt_risk_arm_skip reason=empty_echo` if still mute). - Far-bus listen missing `71` ≠ T340+ silent; without a tap on each Pi↔T340 local drop, swallow vs never-sent is open. Prior pos2/pos5 misses: ACK was sent, trimmed on UART→ETH harvest. - Python-only — no uartd rebuild. Rollback: `2.2.2.24` (mrt-x9-x1-inject-diag). ## 2.2.2.24 — 2026-07-14 — mrt-x9-x1-inject-diag - Focused soak logging for hard-miss vs late local `x1` after eth→UART `x9` (79→71): compact `mrt_x9_inject_diag` (echo/post-DE/peer + arm result), `mrt_x1_after_x9` with `x9_to_x1_ms` + `late_after_miss`, and enriched `x9_no_local_x1` with last inject snapshot. - Python-only — no uartd rebuild. Rollback: `2.2.2.23` (mrt-arm-skip-closed-seq). ## 2.2.2.23 — 2026-07-14 — mrt-arm-skip-closed-seq - Pos1 soak false `x9_no_local_x1` @400 ms: Master poll retry eth→UART `79` re-armed await after synth `0x0D` already completed; copper witness already had local `71`+`0D`. - **`_mrt_risk_arm_after_local_x9`:** skip re-arm when `play_sequence` already closed (`mrt_risk_arm_skip` `seq_already_closed`), when await-0D is active, or when await-x1 already armed for the same seq (do not reset the 400 ms clock). Closing `0x0D` records `_mrt_risk_close_done_seq`. - Python-only — no uartd rebuild. Rollback: `2.2.2.22` (listen-ipc-desync-seq-wrap). ## 2.2.2.22 — 2026-07-14 — listen-ipc-desync-seq-wrap - **Listen crash (pos2 / pos51):** after header commit, non-blocking empty payload `recv` raised `BlockingIOError` → `uartd_ipc_retry` without consuming payload → next poll treated payload as magic → `RuntimeError: bad ipc magic` killed Python. Poll now waits for payload; desync disconnects + `reconnect()`; emit `uartd_ipc_desync` / `uartd_ipc_reconnect`. - **`hat_fixture_listen.sh`:** `LISTEN_AUTO_RESTART=1` (default) restarts Python+uartd on non-zero exit **without** re-wiping capture (`CLEAR_CAPTURE` wipe is first start only). Ctrl-C / SIGTERM do not restart. Tracebacks always append to `fixture_listen.log`. - **MRT regression (2.2.2.21):** `play_sequence` wrap/reset (soak `7→0`) left `latest_seq` high → `status_0d_synth_skip reason=superseded_seq` → `mrt_risk_late_0d_after_x1`. Supersede / synth guard now treat any **changed** seq (including decrease) as newer; same seq still leaves wait alone. - Python-only — no uartd rebuild. Rollback: `2.2.2.21` (local-seq-supersede-close). ## 2.2.2.21 — 2026-07-14 — local-seq-supersede-close - Stuck local close-wait (await `x1` / synth `0x0D` / trunk gate) could leave Contacting Game Server while the bridge process stayed up — new local wager/`x9` did not fully abandon prior close-pending state. - **Local TID only:** for the same `(tid, game)`, any newer `play_sequence` (`> N`, not only `N+1`) supersedes the open wait; missing sequence is latest-wins. Same or lower sequence leaves the wait alone. - Clears await-x1 / await-0D / `x1_pending` / synth-sent / skill bits; closes older/unknown trunk and flushes deferred injects. Emits `local_round_superseded`. Synth skip reason `superseded_seq` if x9 hint seq is stale. - Python-only — no uartd rebuild. Rollback: `2.2.2.20` (listen-ipc-eagain-retry). ## 2.2.2.20 — 2026-07-14 — listen-ipc-eagain-retry - Listen witnesses (pos2 / pos51) died mid-soak with `BlockingIOError: [Errno 11] Resource temporarily unavailable` in `uartd` non-blocking drain: IPC header arrived, payload not ready yet → uncaught EAGAIN killed Python. - **`pog422_uart_client._recv_exact`:** empty-socket EAGAIN still means “no event”; mid-frame EAGAIN waits (select/timeout) for the rest. - **`listen_loop_uartd`:** catch transient IPC/socket errors, emit `uartd_ipc_retry`, continue (no process exit). - Python-only — no uartd rebuild required for this fix. Rollback: `2.2.2.19` (post-de-await-local-x1). ## 2.2.2.19 — 2026-07-14 — post-de-await-local-x1 - Fix pos5 MRT (witness `51` soak): after local eth→UART `x9`, post-DE harvest took only orphan SOH (`0x05`) of the slave `x1` (5 ms + 200 µs idle), then `window_skip split_empty` discarded it and grace skipped the remainder → `x9_no_local_x1` while copper `x1` existed. - **uartd:** post-DE poll **25 ms** (match grace); idle-exit only on complete SOH…EOP frame; lone SOH held for grace/HS glue (`post_de_soh_into_hs_window`). - **bridge/capture:** incomplete post-DE peer → `post_de_soh_hold` (no finalize skip); glue onto next WINDOW. - Rebuild **uartd** on the Pi (`inject-shim`) — Python alone is not enough for the harvest timing. Rollback: `2.2.2.18` (await-x1-hold-rate-limit). ## 2.2.2.18 — 2026-07-13 — await-x1-hold-rate-limit - Fix 2.2.2.17 soak lockup: forever `await_local_x1` deferred every poll → ~80k events / 25 MB rotate spam; eth→UART wedged on D249/9309. - Non-local eth→UART hold after local `x9` is now **2 s** (`MRT_LOCAL_X1_INJECT_HOLD_MS`), not forever; round still stays open for late `x1`→synth `0x0D` until `x1` or new local wager. - Rate-limit `inject_deferred` logs to ≤1/s per reason. Rollback: `2.2.2.17` (await-local-x1-defer-relay). ## 2.2.2.17 — 2026-07-13 — await-local-x1-defer-relay - After local-TID `x9` eth→UART inject, **hold non-local eth→UART injects** until local `x1` is seen (`inject_deferred` reason `await_local_x1`). - No wall-clock abandon: wait clears only on local `x1` (then synth `0x0D` as before) or a **new local wager**. - Matches native wedge robustness: leave the POG422 bus free for the T340+ ACK after we delivered Master `x9` (pos5/51 MRT soak: late copper `x1` was lost under foreign relay inject). Rollback: `2.2.2.16` (listen-uart-busy-diag). ## 2.2.2.16 — 2026-07-13 — listen-uart-busy-diag - When HAT UART is already open, listen/uartd prints actionable busy diagnostics: `fuser`, matching bridge/supervisor processes, `fixture-prod` / `ethergold` unit state, and stop hints. - Clarifies this failure is **port ownership**, not a bad `FIXTURE_POS`. Rollback: `2.2.2.15` (fixture-pos-1-99-witness). ## 2.2.2.15 — 2026-07-13 — fixture-pos-1-99-witness - Widen `FIXTURE_POS` / `--fixture-pos` from seat-only (`1–5` / listen `1–4`) to **`1..99`**. - Allows POG422 copper witnesses without clobbering seat landings (e.g. **`51`** + `PI_LABEL=topbox5w` for pos5 listen-only). - PC archive/normalize discover numeric `logs/N` folders (not hard-coded `1..5`). Rollback: `2.2.2.14` (clear-capture-wipe-dirs). ## 2.2.2.14 — 2026-07-13 — clear-capture-wipe-dirs - `CLEAR_CAPTURE=1` (default bench): wipe **all contents** of `data/` and `logs/` before start (rotated `events.jsonl.*` / `packets.raw.*` included), not only the active allowlist. - Listen matches bridge: shell wipe first, then `--append-capture` so Python does not re-wipe `logs/` after uartd opens `uartd.log`. - Fixes fat Dropbox re-sync of multi-day rotated chunks after each soak. Rollback: `2.2.2.13` (mrt-means-late-0d-after-x1). ## 2.2.2.13 — 2026-07-13 — mrt-means-late-0d-after-x1 - Clarify diagnostics: **true MRT** = local x1 (`0x41`…) without timely closing `0x0D`. Log `mrt_risk_late_0d_after_x1` only for that. - Rename missing-ack watch to `x9_no_local_x1` (rejected/ignored x9 / re-wager) — **not** MRT. - Retest `…_1923-UTC_pos5-keno-mrt-retest-22211`: 14/14 local `0x41` → synth `0x0D` in 29–47 ms; prior `mrt_risk_missing_x1` events were x9-not-accepted, not MRT. Rollback: `2.2.2.12` (early-win-after-tx-lock). ## 2.2.2.12 — 2026-07-13 — early-win-after-tx-lock - Move Phase1/2 early-win inspect (incl. Keno fireworks IPC write) **outside** `_tx_lock`, after pending status `0x0D` synth — so a slow fireworks file write cannot block `41→0D` inject. - Retest archive `2026-07-13_1923-UTC_pos5-keno-mrt-retest-22211`: fireworks was **not** causal for the two `mrt_risk_missing_x1` rounds (both `qualifies=False`); uartd grace absorb is working on good rounds. Remaining MRT = no local `0x41` (game retried `0x40` / bank traffic), not late synth after x1. Rollback: `2.2.2.11` (uartd-post-x9-grace-absorb). ## 2.2.2.11 — 2026-07-13 — uartd-post-x9-grace-absorb - **Keno MRT:** uartd now polls the UART FIFO for 25 ms after eth→UART inject ends with SLV-HS low (port of pyserial `post_x9_inject_grace_hold` / `absorb`). Split/local `0x41` can become a WINDOW so `status_0d_synth` still fires — **does not** synth `0x0D` before local x1. - Python: `mrt_risk_missing_x1` / `mrt_risk_late_0d` after local x9; uartd WINDOW path applies handoff x1 filter + logs `post_x9_inject_grace_absorb`. - BOM: `POG422_EG_VERSION` / uartd+inject → **2.2.2.11** (source-newer rebuild). Driver may stay in `[2.2.2.7, tip]`. Rollback: `2.2.2.10` (component-version-range). ## 2.2.2.10 — 2026-07-13 — component-version-range - Version checks use **`[EG_COMPONENT_MIN_VERSION, EG_COMPONENT_MAX_VERSION]`** (now `[2.2.2.7, 2.2.2.10]`), not identical stamps across Python / uartd / inject / XR driver. - `update.sh` rebuilds inject/driver only when out of range (or FORCE_* / HW_DEBUG / newer sources); IPC lag vs Python tip is a NOTE, not FATAL. - Reverted temporary C/ipc/`MODULE_VERSION` stamp from 2.2.2.9 back to **2.2.2.7** (last real binary change). Python tip may spin without forcing rebuilds. Rollback: `2.2.2.9` (keno-balls-offset-21-raw). ## 2.2.2.9 — 2026-07-13 — keno-balls-offset-21-raw - Keno `0x49` draw balls: **fixed offset 21**, raw display 1–80 (lab + natural). - Removed dual-path / wire−2@19 fallacy; lab `build_keno_jpot_synth` writes raw @21. - Phase 2 TouchEasy decode matches that single layout. - BOM stamp: `pog422_ipc.h` / XR `MODULE_VERSION` / `install_pi.sh` → **2.2.2.9** (clears update.sh mismatch after Python-only spin; superseded by 2.2.2.10 range policy). Rollback: `2.2.2.8` (phase2-toucheasy-keno). ## 2.2.2.8 — 2026-07-13 — phase2-toucheasy-keno - Phase 2 cash early win for **TouchEasy Keno (9309)**: cache `0x40` picks/bet, evaluate `0x49` vs SGVI sheet-10 paytable, local-TID fireworks when win ≥ threshold. - Skip when `progwinval≠0` (Phase 1 owns progressive). - Enable: `FIXTURE_EARLY_WIN_PHASE2_KENO=1` / `sync/enable_early_win_phase2_keno_env.sh`. Rollback: `2.2.2.7` (partial-0e-wire-len-598). ## 2.2.2.7 — 2026-07-13 — partial-0e-wire-len-598 - `COMMAND_WIRE_LEN[0x0E]=598` (Partial Master reply); alt **397** (Jun 19 census mid-EOP). - UARTD C matched. STOM priority includes `0x0E`. - Fixes pos3 `eth_tx_skip wire_len_crop` after Master answered Partial (`…1626-UTC_partial-2226-master-replied`): HS `05,0E…` @598 mis-parsed as `00`@131. Rollback: `2.2.2.6` (partial-04-wire-len-32). ## 2.2.2.6 — 2026-07-13 — partial-04-wire-len-32 - `COMMAND_WIRE_LEN[0x04]=32` (Partial Transfer request); UARTD C matched. - STOM priority/request sets include `0x04` (was documented as `0x02` only). - Fixes pos4 `eth_tx_skip wire_len_crop` on Partial — Master never saw request (session `…1515-UTC_mtc-2225-2ok-partial-fail`). - Full MTC `01`/`09` already OK on 2.2.2.5. Rollback: `2.2.2.5` (mtc-inject-chunk-writesize). ## 2.2.2.5 — 2026-07-13 — mtc-inject-chunk-writesize - `INJECT_TX`: chunk wide USB OUT to `writesize` (DE held across chunks); scale chunk timeout with air time. - Inject lib: `timeout_ms = max(500, air_ms+500)` for large frames. - Bridge: log inject forensics on `uartd_inject` error (`usb_actual`, FIFOs). - Fixes MTC after 2.2.2.4: `0x09`@1170 accepted then `errno=110` ETIMEDOUT (~160 ms) on single 2340-byte URB. Rollback: `2.2.2.4` (mtc-inject-frame-max-2048). ## 2.2.2.4 — 2026-07-13 — mtc-inject-frame-max-2048 - `POG422_INJECT_FRAME_MAX` **256 → 2048** (XR `INJECT_TX` + inject lib check). - Fixes MTC after wire-len parse: `0x09`@1170 reached eth→UART then failed `uartd_inject errno=22` (EINVAL). - Session proof: `…1444-UTC_mtc-2223-still-fail-no-retry`. Rollback: `2.2.2.3` (mtc-09-wire-len-1170). ## 2.2.2.3 — 2026-07-13 — mtc-09-wire-len-1170 - Transfer Master Config: live Master `0x09` HS window is **1170 B** (pos2/3 FIFO `05,09…` + EOP), not legacy `comlengs[9]=1094`. - `COMMAND_WIRE_LEN[0x09]=1170`; alt **1094** retained. UARTD C primary/alts matched. - Fixes MTC reply loss: prior parse fell through to junk `00`@235 → `eth_tx_skip wire_len_crop`. Rollback: `2.2.2.2` (peer-pkt-complete-hooks). ## 2.2.2.2 — 2026-07-12 — peer-pkt-complete-hooks - After eth→UART inject, peeled peer TxD runs **`finalize_handshake_window_for_channel`** (outside TX lock): `packet_complete` hooks + UART→ETH — same as native HS capture. - Fixes pos4 MRT: glued local `x1` (e.g. `61` after `69`) now arms `status_0d_synth` under `status_0d_block_all`. Rollback: `2.2.2.1` (inject-peer-uart-eth). ## 2.2.2.1 — 2026-07-12 — inject-peer-uart-eth - Bridge 101: POG422-origin bytes that land in the **inject RX window** (today only `inject_echo`, then cleared) are peeled after self-echo and **UART→ETH on both ports** — all opcodes, not SCR-only. - Helper `peer_uart_bytes_after_inject()`; event `inject_peer_uart_eth` for bench proof. - Fixes Master `0x0B` (and any peer TxD) swallowed during eth→UART inject so slaves receive the reply. Rollback: `2.2.1.17` (scr-transaction-path). ## 2.2.1.17 — 2026-07-12 — scr-transaction-path - Bridge: **0x0B** (Master SCR enroll reply) joins STOM comm **priority** set — bypasses `game_transaction_gate` and `post_x9_grace_*` inject defer; head-of-line queue like 0x03. - `inject_echo`: **prefix match** scores `match=True` when bus echo = sent frame + trailing bytes (glued SCR + `0x0B` capture artifact). - `inject_echo.jsonl`: new `match_kind` / `post_de_rx_match_kind` (`exact` | `prefix` | `mismatch` | `none`). - BOM stamp: inject/uartd/`MODULE_VERSION` → **2.2.1.17** (driver path unchanged; clears update.sh mismatch). Rollback: `2.2.1.16` (wide-rx-9bit-flag0). ## 2.2.1.16 — 2026-07-12 — wide-rx-9bit-flag0 - UARTD: open port as **CS5 → 9-bit** (Exar quirk) so wide RX flag **bit0 = 9th data bit**, not 8-bit wide “parity error.” - Explains 2.2.1.15 session: native SCR SOH showed `05/00` in `wide_rx.bin` while bank accepted — instrumentation was in 8-bit+wide. - Soft HS flag bit `0x10` unchanged. DTtty still delivers the low 8 data bits only. Rollback: `2.2.1.15` (wide-rx-tee). ## 2.2.1.15 — 2026-07-12 — wide-rx-tee - Driver: parallel **wide RX tee** ring — every preciseflags pair kept as `[data][flags]`; unused RAMCTL bit **`0x10`** = cached SLV-IN/HS (`last_slv_in_active`). - New ioctls: `SET/GET/READ/FLUSH_WIDE_RX` (does not change DTtty protocol path). - UARTD: optional always-on thread (`POG422_WIDE_LOG=1`) drains tee to `${OUTPUT_DIR}/wide_rx.bin` (same wide layout). - Purpose: prove MARK bit / HS state on copper vs 8-bit strip into `packets.raw`. Rollback: `2.2.1.14` (slv-in-sample-restore). ## 2.2.1.14 — 2026-07-12 — slv-in-sample-restore - Inject: restore **cheap** `slv_in_phase_active` samples while DE is high (`after_on`/`body` after bulk, `after_drain` before GPIO7 off). - Fixes false `slv_hs_misalign` storm from 2.2.1.13 (fields left zero after dropping copper_push). - Still no multi-register `copper_push` / FIFO get_reg on the DE-high path (hold stretch). Rollback: `2.2.1.13` (ioctl-de-defer). ## 2.2.1.13 — 2026-07-12 — ioctl-de-defer - `INJECT_TX` MANUAL_DE: assert GPIO7 **inside** the ioctl after mutex/pm, immediately before bulk (userspace no longer asserts first). - Skip FIFO get_reg before/after bulk on MANUAL_DE path (USB controls were stretching DE-high before ioctl returned). - Userspace: hybrid air/TEMT wait from estimated DE assert (`tx_ns`); no copper sampling while DE high. - Goal: `de_hold_us` ≈ air+margin (~6.5 ms SCR) instead of ~17 ms ioctl dwell. Rollback: `2.2.1.12` (hybrid-de-deadline). ## 2.2.1.12 — 2026-07-12 — hybrid-de-deadline - Inject: DE release = **min(TEMT, air_us + post_hold margin)** from GPIO7 assert. - Default margin **400 µs** (`POST_AIR_MARGIN_US`); `pre_tx` stays **0**. - New `de_release_reason=deadline` (4); TEMT/EN_ACT still logged as forensics. - Short WAIT slices (≤5 ms) so air deadline is not overshot. - Goal: shrink Master SCR inject HS/~25 ms `de_hold` toward native ~7 ms / ~6.1 ms air. Rollback: `2.2.1.11` (fall-stamp-clamp). ## 2.2.1.11 — 2026-07-12 — fall-stamp-clamp - UARTD: fall-burst / end-window clamp byte stamps to ≤ HS-fall time (FIFO drain is not post-HS copper). - Analysis note (not a code delay): Master SCR inject `de_hold_us` ≈ **22–30 ms** waiting **TEMT/EN_ACT** — matches pos2 `hs_width_us`. `pre_tx`/`post_hold` remain 0; long “HS after last byte” is DE held for late TX-done, not a post-inject sleep. Rollback: `2.2.1.10` (hs-porch-stream). ## 2.2.1.10 — 2026-07-12 — hs-porch-stream - `HW_DEBUG=1` UARTD: stream-drain RX **while HS is high** (prod still burst-on-fall) so `first_byte` / `last_byte` mono are real. - Events: `hs_porch` + porch fields on `packet_complete` — `front_porch_us`, `byte_span_us`, `back_porch_us`, `hs_width_us`, plus `t_hs_rise/first/last/fall_perf_ns`. Rollback: `2.2.1.9` (listen-session-start-fix). ## 2.2.1.9 — 2026-07-12 — listen-session-start-fix - Fix UARTD listen crash: `session_start` passed `uartd_version=` and `**hat_probe` (duplicate kwarg) → TypeError before any capture (empty `events.jsonl`). - `uartd_common.sh` stop: SIGKILL after ~1s if uartd stuck in USB `skb_wait` so Ctrl-C returns a prompt. Rollback: `2.2.1.8` (uartd-listen). ## 2.2.1.8 — 2026-07-12 — uartd-listen - Restore **UARTD listen** on the flat EG stack: `hat_fixture_listen.sh` / `run_listen_jun11_good.sh` start `pog422_uartd` (inject=0) so pos1/2 witnesses use the same C HS/RX path as bridge, including `HW_DEBUG=1` → `hw_debug.jsonl`. - `uartd_common.sh` passes `HW_DEBUG` through to `make` (bench rebuilds keep the dense timeline). - Prod listen no longer refuses a running uartd — listen owns the daemon. Rollback: `2.2.1.7` (custom-driver-en-act-txdone). ## 2.2.1.7 — 2026-07-11 — custom-driver-en-act-txdone - Fix B1422 `init_xr21b142x_reg_map`: `CUSTOM_DRIVER` (0x060) was written onto the **140x** map — probe skipped bit0=1. Now set on 142x map before GPIO/`CUSTOMIZED_INT`. - TX-done naming: B1422 has no Tx-empty USB IRQ / no LSR. `WAIT_TX_EMPTY` returns `reason` (`already_idle` / `en_act_fall` / …); inject gates DE-off on `temt` only (no fake LSR bit6). - Keep `FLOW_CONTROL` bit3 clear (UART Half-Duplex); `hat_init` clears it if set. - Docs: `CUSTOM_DRIVER` ≠ `CUSTOMIZED_INT`; HS=GPIOB1; GPIO5=EN_ACT sense only. Rollback: `2.2.1.6` (hw-debug-compile). ## 2.2.1.6 — 2026-07-11 — hw-debug-compile - Compile-out `POG422_HW_DEBUG` timeline (bench `make HW_DEBUG=1`; prod default off — no runtime `is_debug`). - Userspace: `pog422_hw_debug.h` ring; uartd HS/RX samples + IPC `HWDBG` → `hw_debug.jsonl`. - Inject: `copper_push` GPIO_STAT/FIFO polls **only** when `HW_DEBUG=1`; denser mid-TEMT copper under debug. - Driver: Table-8 USB INT edge ring (GPIO1 + GPIO5 sense) + `DRAIN_HWDBG` ioctl — GPIO5 **only** from that USB path. - `session_start.hw_debug` from daemon version reply (compile-time capability bit). Rollback: `2.2.1.5` (no-pre-tx). ## 2.2.1.5 — 2026-07-11 — no-pre-tx - Eliminate pre-TX settle: C ignores `opts->pre_tx_us` (GPIO7 assert → immediate `INJECT_TX`). - Launcher/BridgeTiming defaults: `PRE_TX_US=0` (was **3000** on uartd inject — HS→first-byte lead). - Bridge `_inject_pre_tx_us` always returns 0; field kept for log/IPC compat only. - Pair with 2.2.1.4 (no post-hold): DE envelope is assert → bulk → TEMT release only. Rollback: `2.2.1.4` (no-post-hold). ## 2.2.1.4 — 2026-07-11 — no-post-hold - Eliminate post-DE safety delay: C `sleep_us(200)` after GPIO7 release removed. - Launcher/BridgeTiming defaults: `POST_TX_US=0`, `POST_AIR_MARGIN_US=0` (was still 1000/200 on uartd inject path — logged as `post_hold_us=1200` even though TEMT inject ignored it). - Bridge `_inject_post_hold_us` always returns 0; field kept for log/IPC compat only. Rollback: `2.2.1.3` (wire-len-c-sync). ## 2.2.1.3 — 2026-07-11 — wire-len-c-sync - Sync `pog422_uartd.c` capture framing to Python `COMMAND_WIRE_LEN` / `_ALTS`. - Fix stale C `0x20=41` (Python primary is **37**, alt 48); add missing opcodes (0x01/0x09/poker/bingo/…). - HS capture helpers now try longest-first alt lengths (same as `split_packets`). - Does not by itself fix pos2 CMSPAR listen vs inject SCR; that remains a separate RX-stack issue. - **Follow-up:** `install_pi.sh` `EXPECTED_VER` was left at `2.2.1.1` while `MODULE_VERSION` was `2.2.1.3`, so a rebuild could never load the BOM-matched driver (`component_versions_ok=false` on bench). Fixed to `2.2.1.3` — re-run `install_pi.sh` on pos3/4 after stopping bridge. Rollback: `2.2.1.2` (packets-raw-full-hs). ## 2.2.1.2 — 2026-07-11 — packets-raw-full-hs - `packets.raw`: one hex line per HS window = **full copper buffer** (parsed or `window_skip`). - Fixes listen on pre-v16ai builds that only wrote successful `packet_complete` lines (e.g. pos2 `v16ac` lost the 20 B inject fragments). - `window_skip` still carries `hs_window_raw` in events; no duplicate split-packet lines for the same window. Rollback: `2.2.1.1` (skip windows may be missing from `packets.raw` on older listen). ## 2.2.1.1 — 2026-07-11 — temt-gpio-mode-match - Inject `pog422_hat_init` wrote `GPIO_MODE=0x0300`, stomping driver `0x030B` (GPIO5 TX-sense). - Align inject `POG422_GPIO_MODE_MANUAL_SLV` to **`0x030B`** so USB interrupt IN TEMT can fire. - Bench 2026-07-11_2029: 100% `de_release_reason=timeout`, DE ~air+50 ms — LNT activity stayed thick. Rollback: `2.2.1.0` (driver OK; inject still stomps mode). ## 2.2.1.0 — 2026-07-11 — temt-usb-int-tx-sense - `WAIT_TX_EMPTY` blocks on **USB interrupt IN** (Table-8), not FIFO poll. - SLV: GPIO7 stays **manual DE**; GPIO5 = `RS485_EN_ACT` **sense-only** (TX activity). - GPIO5 fall → interrupt URB → confirm `TX_FIFO_COUNT==0` → TEMT → C drops GPIO7. - INT mask `0x03D0` (GPIO1 HS + GPIO5 TX-done). No RS485_EN_ACT on the DE pin. Rollback: `2.2.0.0` (FIFO-poll TEMT). ## 2.2.0.0 — 2026-07-11 — temt-de-rx-harvest - Product BOM `EG_VERSION=2.2.0.0` for bridge, uartd, inject, driver. - Manual DE: `INJECT_TX` skips kernel air `usleep`; C waits TEMT via `WAIT_TX_EMPTY`. - Continuous `poll_fn` harvest during DE-high; post-DE RX window tightened to 5 ms. - Forensics: `de_hold_us`, `lsr`, `de_release_reason` (`temt` / `timeout` / `error`). - Retires air-deadline (`pre+air+post`) DE release. Rollback: `2.1.0.0` (uart-c-spine + inject-gate). ## 2.1.0.0 — 2026-07-11 — uart-c-spine + inject-gate - Product BOM `EG_VERSION=2.1.0.0` for bridge, uartd, inject, driver. - Mode C spine: default `POG422_UARTD=1` + `POG422_UARTD_INJECT=1`; MARK/SPACE Python TX hard-failed on EG2. - C `INJECT` `require_idle` + `idle_wait_ms` — bus-idle wait inside inject burst. - Suspend/resume no longer used for eth→UART inject; post-x9 grace armed from `INJECT_DONE`. - `CHANNEL=both` via two uartd instances (machine + ois sockets). - Wire-ownership doc: DATA+HS LED / RCV INT / `INJECT_TX` are success oracles. Rollback: last pre-EG2 tip was parser `v16cb` + uartd/inject `1.0.60` + driver `1W-pog422-tx-fifo-errno-ch`. ## Legacy map (archives only) | Legacy ID | Notes | |-----------|--------| | `v16cb-raw-bus-framing` / `1.0.60` / `1W-…` | Last pre-2.0 snapshot | | `1Q`…`1W` letter drivers | Wide/preciseflags line; identity replaced by EG four-tuple | | `v16xx` parser soup | Replaced by `EG_VERSION` + optional slug in changelog only |