Consolidate the fw #50 campaign into ≤2 stacked, reviewable PRs per repo (close-out)

[mbreiser]

Close-out task for the fw #50 Mode-3 wedge campaign (2026-09-11/12). Michael's rule: at most two stacked PRs per repo, each explainable on its own — not the 30 (web) + 13 (firmware) commits as they landed under bench pressure. The plan with file lists and the squash strategy is docs/development/mode3-wedge-closeout-plan.md on claude/mode3-wedge-soak.

• webDisplayTools PR A — fail-closed fault lifecycle + post-mortem + soak harness + analyzer (no ring). Squash from claude/mode3-wedge-soak per the plan's file list; pixi run test + format-check green.
• webDisplayTools PR B (stacked on A) — controller telemetry ring, host side (drainer, wire 0xA8/0xA9/0xCC, HEALTH tails, bridge 3.1 rows, reader safety, analyzer columns, reconciliation docs).
• Firmware PR F1 — GET_HEALTH + GET_FIRMWARE_VERSION + soak script (PR Experiment Designer: hover tooltips on filmstrip blocks #53 exists off main; the bench ran the arena-2x10-local cherry-picks — decide base/retarget; document the 4×10-vs-2×10 gotcha).
• Firmware PR F2 (stacked) — ring + diagnostic watchdog. Before opening: collapse GET_HEALTH ver 1→4 to ONE new version; move bench-only watchdog verify fields to a diagnostic opcode or drop; decide the watchdog default for course controllers.
• Update Linear LAB-212 (recovery findings) and LAB-149 (ring status); post the morning analysis on fw 3D viewer pole vector too small and misaligned with expansion pattern #50; add the health-tag/diag-opcode items to fw Experiment Designer: undo/redo for all edits #54.
• Morning analysis + time-boxed stock arena-2x10-local control night (status doc §4.4).

Related: #197, #198, #199, #200, reiserlab/LED-Display_G6_Firmware_Arena#50, #53, #54, LAB-149, LAB-212.

[mbreiser]

Add to the close-out list (found 2026-09-12 23:55 during the watchdog night): (1) Studio.patternSet is rebuilt from the SD listing on every connect, so frame counts for card-only patterns (no manifest entry) are lost after a reconnect → patternFramesByName falls back and the bridge is configured with the wrong frames → out-of-range 0x70 rejects → fault. Proper fix: resolve the frame count from the controller (GET_PATTERN_INFO 0x88) at trial start when the set lacks it. (2) A post-mortem outcome transient in the iteration right after a self-reset should not end the soak. (3) Firmware: free-running refresh timer (SET_FRAME_POSITION must not disarm/re-arm the PIT) — the display starves at command rates near the refresh period AND wedge #5's watchdog-captured PC is the PIT TCTRL store in IntervalTimer::end().

[mbreiser]

From the 2026-09-13 Codex reviews (docs/development/codex-review-2026-09-13-mode3-wedge-fix.md), for the campaign analysis / close-out:

• A/B design: matched exposure blocks (100 Hz primary; 200/286 Hz secondary), same simulator seed/patterns/instrumentation, cold-start vs post-wedge blocks separated, clean hours reported as censored exposure; predeclared stopping rule (0 failures in T h ⇒ 95 % upper bound ≈ 3/T per hour). Mechanism evidence comes from the watchdog context capture (STATE kind 8) and the bench stress reproducer, not from soak statistics.
• Report performance and reliability separately per block: refresh-ISR count, PIT entries, total panel transfers vs changed-frame FRAME records, accepted 0x70s, SD reads, RTT decomposition, telemetry loss.
• Host decode landed in 69c14e6 (0xCB bit 4 freeRunningTimer, ISR ids 4–7, STATE kinds 8/9); wedge-scan.py labels the watchdog context.

[mbreiser]

Close-out addition (2026-09-13 09:55 ET): the fw #50 mechanism comment is posted (reiserlab/LED-Display_G6_Firmware_Arena#50, comment 5653672436). The PJRC upstream report (text B in docs/development/mode3-wedge-upstream-note-2026-09-13.md, reproducer in firmware tools/pit-race-repro/) is held for Frank's review before filing on PaulStoffregen/cores — add it to the close-out checklist.

[mbreiser]

2026-09-13 — performance session additions to the close-out:

• Host branch claude/mode3-perf-sd (stacked on claude/mode3-wedge-soak / feat(studio): controller-fault detection, post-mortem probes, soak driver, GET_HEALTH decode — Studio v0.76 (fw #50) #198): Studio v0.77 — ring-v2 FRAME decode (cf rows gain 3 optional fields), STATE kinds 11–13, 0xCB bit 5, GET_SD_INFO 0xCD → run_metadata.sd_card, js/trial-quality.js (per-trial pass/flagged/unknown, display_gap + trial_quality events, flag never auto-exclude), drainer per-poll budget 40 → 200 chunks, scripts/telemetry-report.py (+ tests, in pixi run test). Belongs in PR B (telemetry) or a PR C stacked on it — Michael decides.
• Firmware branch feat/sd-fastpath-2x10 (local): SD fast path + attribution + 0xCD + hygiene; two Codex diff reviews reconciled; flashed as 3c71953. PR placement: F3 stacked on F2, or folded into F2 — decide with the GET_HEALTH-ver collapse.
• Docs: sd-read-jitter-2026-09-13.md, telemetry-review-handoff-2026-09-13.md (review-session brief: what is logged, three-clock model, round-trip recipe, Teensy-only timing proposals), bench log §10.
• Still open from earlier comments: frame count from GET_PATTERN_INFO 0x88 (the __framesRepatch hook is still needed), transient-after-self-reset soak rule, rig chip shows "bench ≠ protocol" even with an explicit matching rig (cosmetic, seen today), test_crashreport_passthrough fails after a HalfKay reflash (junk RAM region), test_overfill_evicts_oldest… timing-flaky.

[mbreiser]

2026-09-13 12:30 ET: host branch claude/mode3-perf-sd pushed (Studio v0.77; stacked on #198). Bench: two iterations on fw 3c71953 = 0 stalls / 40 trials pass vs 4 clusters per iteration the night before — see reiserlab/LED-Display_G6_Firmware_Arena#54 (latest comment). Firmware branch feat/sd-fastpath-2x10 is local (tip f6c11d2 built, not flashed; needs a Codex diff review); 3c71953 is on the controller. Causal reflash test deferred.