A second round of board performance work, on top of #813. Each change is its own commit:
- Layout tick: each agent's work group, group node and bearing are cached for the tick, instead of looked up per pair, and the work-node grid uses typed arrays.
- Raster and zoomed-in drawing: collapse membership is recomputed only for the agents an event names, plus each person's top-level agents, which also depend on the person's host count. The widest-zoom theme background is a cached layer, reused until the camera or the groups change.
- First load: history and the tally avoid per-record allocation, so applying
/viewat 10,000 agents takes 81 ms instead of 428 ms. - Mock clustering: on the mock page it runs in a
web/work-worker.jsWeb Worker, and falls back to the main thread where there is noWorker. The live board takes its groups from the index. This addsworker-src 'self'to the CSP ininfra/files/Caddyfile;script-src 'self'already allowed the same-origin worker. It ships with the next deploy. - Benchmark:
board.mjsgains--dpr, long-task counts and first-load time.
Frame time at 10,000 agents from web/dev/bench/board.mjs, in headless Chromium with software GL, main and this branch run back to back on a busy machine (load average 17-21):
| zoom | p50 main → this | p95 main → this |
|---|---|---|
| whole board | 78.7 → 57.7 ms | 93.6 → 105.4 ms |
| one household | 130.7 → 94.1 ms | 145.6 → 117.7 ms |
| close-up | 159.2 → 85.9 ms | 195.1 → 120.4 ms |
The whole board's p95 rose; on a machine this loaded that may be noise. What remains is mostly layout.js:tick, with partOfAgent at 8-13% when zoomed in.
All web tests pass apart from the two known failures, the shells test and the placeholder guard. New tests cover the worker's fallback and protocol. A runtime check at 2,000 agents compared every cached collapse answer with a fresh computation over 10 s of live events and found no mismatch. cargo test -p vibescrobble-query --test caddy passes with the new worker file.
🤖 Generated with Claude Code