perf(app): stop the per-second shell re-render, memoize CLI resolution, halve renderer assets

[ozymandiashh]

Three quiet desktop inefficiencies, each measured before and after, each pinned by a regression test.

1. The shell stops re-rendering once a second

AppMain owned a wall-clock setInterval(1000) whose only consumer was the footer's "refreshed Ns ago" label, but every tick reconciled the whole tree — sidebar, hero, daily chart, activity heatmap, every table — 60 times a minute whether or not any data had changed.

The per-second tick now lives in a leaf RefreshedAt component that owns just that label. The shell re-renders on real state changes plus a 15-second day check that fires exactly when the local calendar rolls over; that rollover matters because the overview memo keys bake in a today/month boundary, so midnight must still produce one re-render to keep "Today" honest.

App.renderChurn.test.tsx pins both properties: on the base the shell re-rendered on every clock tick and on same-day checks, on this branch neither.

2. CLI resolution stops re-scanning the filesystem per request

Every read (each section poll, each prefetch warm) resolved the codeburn binary from scratch: a stat sweep over every PATH entry plus a readdir of the nvm versions tree — per request, purely to return the same path.

Resolution and the derived spawn PATH are memoized against the full set of inputs that can change the answer (CODEBURN_BIN, CODEBURN_BUNDLED_CLI, the dev-server and persisted-path overrides, CODEBURN_PATH_DIRS, PATH, NVM_DIR), so an unchanged environment costs a string compare. Safety valves:

• a miss is never cached, so a CLI installed — or a path persisted via locate-CLI — while the app is open is still discovered on the next read;
• a spawn error drops the memo, so a deleted binary cannot pin the app to a dead path (covered by a test that deletes the resolved binary and watches the next request find the successor);
• __resetCliResolutionForTests() keeps one test's filesystem layout out of the next.

Measured on this machine (18 PATH entries): 26.5 µs → 2.6 µs per resolution, and the syscalls leave the steady-state path entirely.

3. Renderer assets halve

The brand mark was an 880×880 713 kB PNG (3.1 MB RGBA decode) rendered into 20–76 px boxes, so every mount decoded megabytes to draw a 20 px sidebar logo; the cold-start splash shipped a 2.6 MB VP9 clip.

• flame-mark.png: a 192 px lanczos cut of the same art (53 kB, −93%), verified side-by-side against the original at all three rendered sizes (20/52/76 px).
• splash-loader.webm: re-encoded at VP9 CRF 30 — 1.29 MB, −50%, SSIM 0.994 against the source; video-only file, no audio track affected.

dist/renderer: 4.0 MB → 2.2 MB. The JS bundle is untouched.

Verification

• tsc --noEmit clean
• full app/ suite green on this branch rebased onto current main: 997 passed, 4 skipped
• npm run build green
• churn tests verified to fail on the pre-change code and pass after
• deliberately out of scope: bundle code-splitting (gsap ~50 kB of the 510 kB chunk) — separate PR material

Changed

• The desktop shell stops re-rendering the whole tree once a second. AppMain owned a wall-clock setInterval(1000) that existed only so the footer could print "refreshed Ns ago", but every tick re-rendered the sidebar, the hero, the daily chart, the heatmap and every table 60 times a minute whether or not any data had changed. The per-second tick now lives in a leaf RefreshedAt component that owns just that label, and the shell itself re-renders only on real state changes plus a 15-second day check that fires exactly when the local calendar rolls over. Renderer render-count regressions are pinned by a churn test.

Fixed

• The desktop app stops re-scanning the filesystem for the CLI on every request. Each of the app's JSON reads resolved the codeburn binary from scratch: a stat sweep over every PATH entry plus a readdir of the nvm versions tree, repeated per request. Resolution and the derived spawn PATH are now memoized against the full set of inputs that can change the answer; a miss is never cached, so a CLI installed while the app is open is still discovered, and a spawn error drops the memo.
• The renderer ships a 53 kB flame, not a 712 kB one, and half the splash video weight. A 192 px lanczos cut of the same art replaces the 880 px original and the cold-start splash video is re-encoded at VP9 CRF 30 (SSIM 0.994), together cutting the renderer's bundled assets roughly in half.