HW CI: back to one capture per build; keep measure.py --runs

[bwhitman]

Follow-up to #897. The 3× averaging trial (re-bench of #896) answered the noise question: run-to-run spread within one flashed binary is only ±1–2 µs (PR: 3704/3702/3703 · main: 3674/3672/3672), while the ~±30 µs full-chord delta on that no-op PR showed up identically in every run. The noise is per-binary (code/flash-cache layout jitter between builds), so averaging board resets can't reduce it — it just costs ~80 s per bench job.

• amy-hwci.yml: drop --runs 3 from both measure invocations, restore the single-run comment text, and note the finding in the header comment. The per-run artifact globs stay (harmless, and correct if --runs is ever re-enabled).
• measure.py / hwci_report.py: unchanged — full --runs support (per-run stats, cross-run means, spread line in the report) stays for ad-hoc bench investigation.
• README: records the finding so nobody re-runs this experiment from scratch.

🤖 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. Each build is flashed once and captured 3 times (board reset between runs); the numbers are per-run means.

✅ PASS — the bench ran the test to completion.

notes held	main @ 555c0a6	this PR	Δ
1	1603	1606	+3
2	1799	1802	+3
3	2481	2485	+4
4	2720	2720	+0
5	3464	3468	+4
6	3673	3676	+3

Full chord settled render μs: 3676 (was 3672, Δ +0.1%) (peak 3686, 117 samples, mean of 3 runs)

_{per-run full-chord μs — this PR: 3678 / 3675 / 3676 · main @ 555c0a6: 3673 / 3672 / 3672}

⬇️ 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/.}