feat: golden render baselines, recorded flake recheck, sound-library rebuild docs

[jamiesun]

What

P0 verification stability (profile check):

• Every passing patch now records a render_sha256 golden baseline; diffing two certification reports pinpoints exactly which patches drifted after a library/tool upgrade.
• A first-pass render mismatch no longer fails blindly: full diagnostics are captured (load average, sfizz identity, per-render hashes/timings, RMS difference ratio), then one isolated recheck runs. Recheck pass → load_sensitive_flake warning with evidence retained; recheck fail → hard failure with both snapshots.
• Root cause: on a loaded shared machine, sfizz background streaming raced and produced false "two renders differ" verdicts; single renders were always byte-identical.

Sound-library explainability (docs):

• New docs-site/src/sound-library.md: the sample corpus is never published (some upstream licenses forbid repackaging), but the full recipe is — acquisition channels for all 24 libraries with versions and licenses, directory/manifest contract, upstream-defect repair convention, certification workflow, and a minimal third-party rebuild walkthrough.
• Roadmap standing program gains principle 6 ("explainable and rebuildable") plus adjudicated sections: remaining-gap disposal, repair convention, P0 shipped, selection-governance direction.

Validation

• 94/94 tests green (3 new E2E: golden hash assertion, flake recovery via counted fake sfizz_render, persistent-failure path); fmt + clippy clean.
• Real-corpus certification: 101 mappings / 85 patches, 0 failures, all with golden hashes.
• mdbook build docs-site passes with the new page.

Acceptance matrix

M7 row updated for the recorded-recheck semantics; failure path covered by the persistent-failure E2E test.