Closed loop: one out-of-range 0x70 (status 1 + error glyph) at the start of each trial whose frame count differs from the previous — bridge frame modulus applied after the first frame

[mbreiser]

Bench 2026-09-11, first fw #50 soak iteration (Studio v0.76, sim at 100 Hz, soak_mode3_closed_loop.yaml with a 200-frame and a 20-frame pattern alternating): 110,641 of 110,643 frame commands OK, 2 rejected — both the first 0x70 of a cl_grating_20f (20-frame) trial, carrying an index computed with the previous trial's modulus (156 and 115 → CE_BAD_PARAM "index out of range", which also flashes the error glyph on the arena).

Cause: the runner's fictracApply op does bridge.setConfig({ frames }) then setApply(true); the bridge (Python) applies the config asynchronously, and the first frame message after setApply was computed with the old frames. It is the one-frame tail of the frame-modulus fix (cd6637e).

Options:

1. fictracApply waits for the bridge's cfg ack (the [cfg] applied … message the bridge already prints) before setApply(true).
2. The bridge client clamps locally: setConfig({frames}) also sets _clampFrame = i => ((i % frames) + frames) % frames (the heading→index map wraps, so a stale index modulo the new count is the right frame when the old count is a multiple of the new one, and at worst one wrong frame instead of a reject + glyph).
3. Both.

Also seen in the same run, for the record: 203 of 110,664 replies (0.18 %) took 50–135 ms (host-observed, incl. queue time), not clustered at step boundaries — unexplained until GET_HEALTH (fw #53) can say whether the controller or the browser paused.

Analyzer table: pixi run python scripts/wedge-scan.py soak-logs/arena-log-20260911-230212-393.jsonl (bench01 log, kept locally).

[mbreiser]

Fixed on the #198 branch (16bb3ca, option 2): FicTracBridgeClient._drain wraps the pending index into the configured frames modulus before the consumer clamp (((idx mod n) + n) mod n; no configured count → untouched), so the one late-modulus index per trial change lands on the right frame instead of being rejected with an error glyph. This also matters because controller rejects now count toward the fault window. Test added (156 → 16 for a 20-frame pattern; passthrough with 200; negative wraps). Will close when #198 merges.