Brummel
6933dc4c88
Single-iter milestone, single-artefact data update. Architect:
CLEAN — no design-ledger drift, no contract touched, no Rust source
touched. The `verification` contract (design/INDEX.md line 100)
ratifying-test (`bench/architect_sweeps.sh`) exits 0 on HEAD; the
script itself was not modified, only its input data
(`bench/baseline.json`). `jq`-keys-grep confirms zero
`max_us`/`p99_9_us` residue, matching the spec's acceptance check.
The `note` field reads as honest current-state with rationale
pointer to `docs/specs/2026-05-20-bench-harness-recalibration.md`
— no JOURNAL refs, no Wunschdenken anchors.
Bench: 0/0/0 after one stochastic-spike clarification.
- `bench/check.py` exit 0 — 57 metrics, 0 regressed (post-recapture).
- `bench/compile_check.py` exit 0 — clean.
- `bench/cross_lang.py` first run exit 1 with one borderline
REGRESSION (`bench_compute_intsum.ail_rc_s` +37.94% vs 35% tol on
the sub-millisecond intsum workload — 0.5ms total, dominated by
syscall/IO startup not algorithmic work). Bencher-pattern re-run:
exit 0, same metric at +14.90% (23-percentage-point swing across
two byte-identical-binary runs = classic single-sample jitter on
the smallest workload, not drift). No baseline change to
`bench/baseline_cross_lang.json`; the existing tolerances are
fine.
Resolution: carry-on. No tidy iteration. No baseline change in
this audit-close commit (the iter commit `21cfccf` already wrote
the `bench/baseline.json` recapture per the iter's `--update-
baseline` step). No --update-baseline run on `cross_lang.py` (the
intsum flicker is jitter, not drift; the second-run clean exit
proves the baselines hold).
Architect drift item, [low], commit-prose correction:
The iter commit `21cfccf`'s "Follow-up noted" paragraph claims
`skills/implement/agents/ailang-implement-orchestrator.md`
references a Phase-4 per-iter journal write. That claim is wrong.
`grep -n "journal" skills/implement/agents/ailang-implement-
orchestrator.md` returns zero hits; Phase 4 of the orchestrator-
agent is `BLOCKED.md` (line 214: `### Phase 4 — On PARTIAL/BLOCKED,
write BLOCKED.md`). The orchestrator-agent's own end-report carried
the same factual error into the iter commit body. Forward-fix
recorded here: there is no orchestrator-agent staleness to chase;
no follow-up tidy iter is needed for this. main HEAD is sacrosanct
— the 21cfccf commit body stays as-is on disk, the correction
lives here.
Pipeline this milestone followed:
brainstorm (PASS, trivial-spec path — no compiler/checker/codegen
claims)
-> planner (recon + plan + self-review 8/8)
-> implement (DONE 4/4, 0 re-loops across all tasks)
-> audit (CLEAN, no tidy, one [low] commit-prose forward-fix)
Closing Gitea #15 + #16. The `closes #15` / `closes #16` trailers
live on the iter commit `21cfccf` and will fire on push.
Spec: docs/specs/2026-05-20-bench-harness-recalibration.md
Plan: docs/plans/bench-harness-recalibration.1.md