dangernoodle-io / TaipanMiner / 26697766720

90%

push

github

fix(mining): adaptive bucketed bench with steady-state convergence detection (#461)

- bench helpers now run iters in 200-iter buckets, declare "settled" when 3 consecutive buckets are within 5% of each other, and report us_per_op/khs from the settled portion only
- replaces pr #460's fixed 200-iter warmup which was empirically inadequate (s2 takes ~4900 iters to settle, settling profile differs per board)
- new sha_bench_result_t fields: settled (bool), settled_after_iters, settled_iters, settled_total_us — surfaced in /api/diag/benchmark response
- boot microbench bumped to 10000 iters (~7 ms total, negligible) so settling + 3 settled buckets always fit
- fallback path when iters too small / never settles: report full-window average + settled=false flag + warning log
- self-adapting; future hardware with different settling profile works without code changes
- closes the bench-vs-ceiling-spread regression observed across the fleet post-pr #459/#460

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>

2595 of 3258 branches covered (79.65%)

Branch coverage included in aggregate %.

4 of 4 new or added lines in 1 file covered. (100.0%)

5252 of 5426 relevant lines covered (96.79%)

152203.72 hits per line