dangernoodle-io / TaipanMiner / 26698603066 / 2

90%

main: 90%

push

github

fix(mining): min-of-tail steady-state estimator + 2% convergence threshold (#462)

- pr #461's mean-of-settled-portion estimator over-reported us_per_op on monotonically-decaying ramps; wroom32 reported ceiling=314 while live mining sustained 360 — impossible (mining > ceiling)
- root: 5% threshold trips "settled" on slow ramps (each step <5% but still decaying); mean across settled portion includes ramp residue
- fix: tighten BENCH_SETTLE_THRESHOLD_PCT 5 → 2, AND use min-of-tail-buckets instead of mean as the steady-state estimator (min bucket is closest to true steady-state regardless of ramp shape)
- applied to both sha256_hw_bench_pass2 (ahb) and sha256_hw_dport_bench_pass2 (dport)
- adds diagnostic log line with mean/min ratio so future ramp profiles surface — ratio >1.05 means residual ramp the old estimator would have missed
- host tests updated for new settled_total_us semantic (single min bucket, not summed)
- on-device validation (wroom32 372.16.1.81, tdongle 172.16.1.78, c3 172.16.1.110, s2 172.16.1.107): ceiling > live mining on all 4 boards (was failing wroom32 + s2); bench-vs-ceiling spread <0.01%; wroom32 mean/min ratio at boot = 1.04 (empirical proof old estimator over-reported by 4%)

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

848 of 1045 branches covered (81.15%)

Branch coverage included in aggregate %.

1978 of 2090 relevant lines covered (94.64%)

395092.13 hits per line