chore(bench): record why BASELINE_ANCHOR_HZ must not be captured on GHA

[johnleider]

What

Comment-and-warning-text only. BASELINE_ANCHOR_HZ stays null, computeScale still returns 1, no artifact changes.

Why

The doc comment on BASELINE_ANCHOR_HZ currently reads:

To capture: run metrics-regen, read apparatus.anchors out of the produced benchmarks.json, paste it here…

Followed literally, that arms v0's benchmark unit off a GitHub Actions runner — which .claude/specs/2026-07-24-dedicated-bench-host.md §6 had already ruled out three days ago ("the in-flight GHA median-of-three capture was halted rather than committed"). The instruction and the decision disagree, and the instruction is the one a maintainer reads at the moment of temptation. metrics:check's warning had the same gap — it named the constant but not the host.

Evidence the spec establishes, now cited at the constant:

• runs-on: ubuntu-24.04 pins the image, not the CPU; the pool spans ~1.9x single-thread throughput.
• At least two vendors, and they do not differ by a scalar: 0.55–0.86 per-anchor Xeon 6973P-C → EPYC 7763. One scale factor cannot describe that.
• A median of three GHA runs does not help — with two vendors in three draws, the median selects the majority vendor, not a typical host.

Measured here on the committed artifacts, GHA also fails the spec's own acceptance bar (±2% aggregate, ±5% worst anchor) against itself — same job, same host, same runner image:

comparison	trimmed gmean	per-anchor spread	worst anchor
1.0.0.json (runs:1) vs benchmarks.json (runs:3)	0.9909	0.916–1.015	a06 map build 0.916
1.0.1.json (runs:1) vs benchmarks.json (runs:3)	0.9151	0.776–1.016	a09 string build 0.776
1.0.1 vs 1.0.0 (both runs:1)	0.9274	0.805–1.006	a12 grouped aggregate 0.805

Part of that is 1-run vs 3-run median noise rather than host disagreement — but metrics:history snapshots are runs:1, so it is the noise the baseline would actually be asked to correct.

Notes

• No changeset: scripts/ only, and changeset-reminder.js gates on packages/*/src/. Ships no package version → master train, chore type.
• Verified: node scripts/check-benchmark-artifacts.ts exits 0 and emits the amended warning; eslint clean on both files.

[johnleider]

Obsolete. This documents why BASELINE_ANCHOR_HZ must not be captured on a GHA runner; #749 deleted BASELINE_ANCHOR_HZ, computeScale, and scripts/lib/calibration.ts entirely, because measurement showed the normalization made results worse rather than better (per-bench disagreement on identical code: 1.81% raw vs 3.53% normalized).

The concern behind this PR was right and is now enforced structurally instead of by comment: benchmarks are measured only on the fixed reference host, CI fails if it touches benchmarks.json, and the reasoning is recorded under "Why there is no host calibration" in .claude/rules/benchmarks.md.