HW CI: average 3 capture runs per build to cut bench noise

[bwhitman]

Back-to-back HW CI runs of effectively no-op PRs have been showing full-chord render µs deltas of ~±33 (e.g. #896), which is bench noise, not the PR's cost. This makes each bench measurement the mean of 3 runs: flash once, capture 20 s, reset the board, capture again — 3× — for both the PR build and its merge-base baseline.

Changes

• measure.py --runs N: flashes once, then resets and re-captures N times (LoadTestChord restarts its chord sequence on reset, and capture() already pulses RTS at t=0, so no re-flash is needed). Run 1 keeps load.csv/serial.log; runs 2+ add loadK.csv/serialK.log. meta.json gains a per-run runs list, and the top-level render_us_final/render_us_max/n_samples become the mean/max/total over the clean runs. A run that gets doubly hit by external bench interference keeps its logs but is dropped from the averages (and still FAILs the verdict, as before). Default is --runs 1, so make speedtest and the sweep are unchanged.

• hwci_report.py: averages each chord size's settled load across the clean runs, prefixes per-run problems with run N:, and adds a per-run full-chord line to the PR comment so run-to-run spread is visible at a glance:

  Full chord settled render µs: 1610 (was 1555, Δ +3.5%) (peak 1640, 117 samples, mean of 3 runs)

  _{per-run full-chord µs — this PR: 1600 / 1600 / 1630 · main @ abc1234: 1550 / 1560 / 1555}

  Old single-run meta.json files (no runs key) still analyze, so a dir from an old sweep or a mixed-version bench run doesn't break the report.

• amy-hwci.yml: passes --runs 3 for both builds, uploads the per-run logs/traces (serial*.log / load*.csv globs), and updates the comment text. Adds ~80 s per bench job — well inside the 20 min timeout.

Testing

Simulated the whole pipeline with upload/capture stubbed to synthetic UART lines: 3 clean runs (means + spread line verified against per-run stats), an interfered run 2 (excluded from means, run 2: errors, FAIL verdict), single-run mode (byte-identical report shape to today), and a legacy no-runs baseline dir. All 15 checks pass; workflow YAML validated.

Note: the bench checks out harness scripts from main, so this PR's own HW CI comment will still be a single-run one — the 3× averaging takes effect for PRs benched after this merges.

🤖 Generated with Claude Code

[github-actions]

🎛️ AMY HW CI (AMYboard bench)

Flashed this PR's AMY (LoadTestChord: 6-voice Juno patch=1, one held note every 2 s) onto the physical AMYboard and measured the smoothed render load as the chord grows — back-to-back with the same sketch built at the PR's merge base, so Δ is this PR's own cost.

✅ PASS — the bench ran the test to completion.

notes held	main @ 73da71e	this PR	Δ
1	1603	1605	+2
2	1800	1804	+4
3	2479	2489	+10
4	2720	2720	+0
5	3465	3466	+1
6	3674	3677	+3

Full chord settled render μs: 3678 (was 3672, Δ +0.2%) (peak 3686, 39 samples)

⬇️ Artifacts: serial log · load trace · report

_{Self-hosted bench (amyboardci). FAIL means only that the test could not run — the load values are informational, with no threshold and no audio compare. See tools/arduino_loadsweep/.}