Bind each snapshot's profiles page to its own freeze

[DavidBakerEffendi]

analyzerProfiles() read currentSnapshot from module scope and AnalyzerProfiles.astro took no snapshot prop, so every snapshots/<slug>/profiles/ page rendered the current freeze's figures rather than its own.

After #152 this became visible in the built site: dist/snapshots/v0-6-0/profiles/ and dist/snapshots/v0-6-1/profiles/ carried v0.7.0's numbers — Bifrost 225 of 369 positive assertions on all three pages, where v0.6.1 published 223 and v0.6.0 published 219. That contradicts the promise snapshots/index.mdx makes, that "no page below has been rewritten since its freeze".

The fix

• docs/src/data/profiles.ts — analyzerProfiles(snapshot: Snapshot = currentSnapshot) takes the snapshot as a parameter and derives both coreKernelPopulations(snapshot.results) and kernelCorpusDistributions(snapshot) from it. The default keeps the unversioned call site working.
• docs/src/components/AnalyzerProfiles.astro — accepts a snapshotVersion prop resolved through snapshotByVersion, in the shape LatencyTier.astro already carries.
• The three profiles.mdx pages name their own version.
• The component's "every latency caveat on the latency page" link now resolves to the page's own freeze rather than the newest one — the same leak in link form, which sent a reader of the archived v0.6.0 page to v0.7.0's caveats.

The latency axis is deliberately untouched

profiles.ts reaches latency through kernelCorpusDistributions, which selects its corpus by snapshot.latencyEvidenceRelease. v0.6.0, v0.6.1 and v0.7.0 all bind the v0.6.0 corpus, so passing each page its own snapshot leaves all eight speed marks identical across the three — which is exactly the mixed-run caveat v0.6.1's and v0.7.0's pages already carry, now derived rather than coincidental.

Verification

npm test 7 pass · astro check 0 errors, 0 warnings, 0 hints · 52 pages built. Read back out of the built HTML, the three pages now differ exactly where they should and nowhere else:

	v0.6.0	v0.6.1	v0.7.0
Bifrost recall	59.3% (219/369)	60.4% (223/369)	61.0% (225/369)
Bifrost precision	98.2%	100.0%	100.0%
other seven analyzers' axes	identical	identical	identical
eight speed medians	identical	identical	identical

Bifrost alone moving is what both caution blocks assert — v0.6.1's "the other seven analyzers did not move at all between these freezes", and v0.7.0's "three of the four re-pins moved nothing at all on the correctness axes".

Not an amendment, on my reading — flagging for the maintainer's call

No frozen evidence, results model, manifest, or scored cell changes here. The correct figures were already in each snapshot's own results model and the site was reading the wrong one, so this is a rendering defect rather than an evidentiary correction under the docs/new-analyzer.md taxonomy — no A<n> number taken. It did, however, put wrong figures on two published pages for the life of #152, so if that reads as warranting a dated amendment, say so and I will add one.

No regression test

docs/src/data/*.test.ts runs under bare node --test; profiles.ts → snapshots.ts pulls in extensionless specifiers and attribute-less JSON imports that only Vite resolves, so the module cannot be loaded there without new machinery. The built-HTML comparison above is the check.

🤖 Generated with Claude Code