--- name: museum-thumb-render-nondeterminism description: "FIXED 2026-07-20: thumbnail re-render nondeterminism was CSS-animation timing drift, not PNG encoder jitter as first guessed — fixed by never re-rendering unchanged prototypes" metadata: node_type: memory type: project originSessionId: 1c45c406-4f75-4a39-8893-a98aca437eb7 modified: 2026-07-21T02:58:52.705Z --- FIXED (2026-07-20, commit 43390e6, folded into the path-based-naming + redirect-ledger session). Originally recorded as: fresh/bare clones cold-render all 28 Playwright prototype thumbnails, and ~11-13 of `museum/site/assets/thumbs/*.png` come out with a few dozen-to-hundred differing bytes each, a DIFFERENT subset every time — diagnosed then as "PNG re-encode jitter," unfixed pending a real cause. **That diagnosis was wrong, and the correction matters.** 27 of the 28 prototypes carry `@keyframes`/`animation:` rules, several genuinely infinite (`coreSpin 9s linear infinite`, `corePulse 2.6s ease-in-out infinite`, a pulsing status lamp). `generate.mjs`'s capture waits a FIXED 600ms after page load with no synchronization to animation phase, so ordinary process-scheduling jitter between runs lands the screenshot at a different rotation angle / glow intensity each time — REAL pixel content drift, not lossless-encoder nondeterminism on identical pixels. This means the other candidate fix the previous session named ("force deterministic PNG encoding") would not have touched the actual cause at all — the encoder was never the problem. **The fix doesn't need to fully solve animation timing to work.** `screenshotPrototypes` now checks the already-committed `museum/site/assets/thumbs/<sha>.png` (the PREVIOUS successful build's output — tracked, not gitignored) BEFORE the local `.cache/thumbs` and before ever launching Playwright. A clone regenerating an unchanged manifest finds every thumbnail already sitting in the committed site and re-renders nothing, so the question of what exactly makes Playwright's capture non-reproducible becomes moot — rendering, and therefore the opportunity for drift, only ever happens once per distinct prototype content, ever, matching the "measured once, frozen" treatment `release-measurements.json` already gets. This was deliberately NOT "commit `museum/.cache/thumbs` too" (the other named candidate) — that would duplicate, byte for byte, storage the site already carries at `museum/site/assets/thumbs`. Verified via fresh GitHub clone AND a bare-clone-derived clone (both with an empty, untouched gitignored cache): `prototypes: 28 total, 28 cached, 0 to render` in both, `git status` clean, stamp `80bf9bb · f64fb7397980c9cb` matching the local run and each other. See [[museum-path-based-naming-and-redirects]] for the session this was folded into, and [[museum-lf-normalization]] / [[museum-readme-txt-leak]] for the earlier determinism fixes this one follows the same pattern as.