[tap]

What this changes

PR 1 of the comparison plan (the rest: competitor set, bridge, rational).

• Freshness gate for the comparison. compare.yml now harvests its counts with provenance (tree SHA, DspTap pin, run ID, toolchains) into async/bench/compare_counts.json (scripts/harvest_compare.py) and uploads it. scripts/update_compare_docs.py derives COMPARISON.md's embedded tables and key figures from that record, and a new compare-docs CI job fails on drift, the same shape as icount-docs.
• Async refresh. Seeded from today's compare.yml run 37920542901 on fc2949e (DspTap ef6fdc4). Every cell reproduces the 2026-09-25 tables except the one the Helium Q15 dot moved: M55 Q15 steady state, 1,163 → 861. Footnote ² and the construction paragraph are reworded around the generated figures; a doubled parenthetical is gone.
• Construct-only ratchet legs, after rational's: async construct_q15, bridge construct_{up,down}_q15_eco. Each is the Q15 workload with its stream removed, the converter built and the same input fixture synthesized, then one block processed, so stream − construct is the loop alone. Baselines recorded on all three targets; README tables regenerated.

Why

The SampleRateTap rows in COMPARISON.md were hand-typed under a date line, so a DspTap bump that moved a count failed nothing: the Helium dot landed on 2026-09-30 and the doc stayed at the September numbers until this morning's run exposed it.

The construct legs are the mitigation PERFORMANCE.md has queued since July and bridge's M7e asked for. On the M33 the Q15 totals are mostly constructor (async 89 %, bridge 78 %), so a hot-path regression there was diluted several times over before the ±3 % gate saw it.

Verification

• Comparison numbers: CI-measured (run 21 on main), not local. The derivation matches the hand tables cell for cell at the 2026-09-25 pin; update_compare_docs.py is idempotent.
• Ratchet: built with CI's configure line and the tree's plugin on M55, M33 and Hexagon (pinned toolchains, plugin-enabled qemu-hexagon 8.2.2). Every existing baseline is identical to the instruction before and after; only the new rows are added. The PR's ratchet jobs passed on all three targets, so the new rows match CI's runners.
• Cross-check worth having: pipeline_q15 − construct_q15 on M33 is 1,188/frame against the 2 s/4 s split's 1,138. The 50 is the servo's one-time lock-in, which the split counts as setup; both are measured and PERFORMANCE.md says which answers which question.
• Why the fixture is in the construct leg: the pipeline synthesizes 0.25 s of input with libm sin(), 52 M soft-double instructions on M33. Without it in the subtrahend, "stream − construct" was 1,729/frame, i.e. not the stream. Rational's construct_q15_eco has that gap; a follow-up could align it.
• After The 2026-10 adversarial audit and its response: S0 to S5, family version 0.6.0 #72 merged: main is merged in. On the merged tree: no conflicts, the new retired-words gate reports nothing, both freshness gates regenerate with no diff, gcc -Werror with every engine's WERROR and the C ABIs on, 305/305 tests, book clean.
• Host: clang-format and clang-tidy clean on the changed units (bridge's icount TU is under the tidy gate; async's is excluded per style.yml). Both workflows parse.
• Not run: the new compare.yml harvest step end to end on CI (it runs on dispatch only). The harvest script was exercised on run 21's log, which is the same input the step writes.

Notes for the reviewer

• New file async/bench/compare_counts.json is the comparison's equivalent of baselines.json: a measured record with provenance. Re-measure by dispatching compare.yml, downloading the artifact, committing it and the regenerated block.
• Provenance pin vs. tree pin. The record names DspTap ef6fdc4; after The 2026-10 adversarial audit and its response: S0 to S5, family version 0.6.0 #72 the tree pins d7ebe4b (the chain.h fix, rational only). The counts stand — nothing async or bridge compiles changed — and the gate deliberately does not require pin equality, so a bump that moves no kernel does not force a 10-minute dispatch. The provenance line says what was measured; the reader decides whether the pin since matters.
• The 12-channel pipeline has no construct leg (different FIFO and fixture sizes); its gate keeps the old dilution. PERFORMANCE.md says so.
• Docs that still cite the 2 s/4 s steady state (1,138 / 3,326 in the README, book and Pico READMEs) remain correct; they describe the split method, which stays.
• Host throughput table untouched, labelled indicative, as decided.

🤖 Generated with Claude Code

https://claude.ai/code/session_01G3HxzEGiZp7jStoMuYNuBK