Intent: warm home holds p95 under the stated budget in test/home-cache.test.ts goes red in pulls that never touch home (pull #5079, a two-line docs/DESIGN.md change, check run 01m5hc81ck9rrt6e7qhs33wtdd: p95 698 ms against 200 ms, and still red when the shard gate re-ran the file alone), so good work is parked.
What was measured (bench fixture, sandboxed crew machine, timeRoute warm, 20 requests): owner home p50 ≈ 95 ms, p95 ≈ 200 ms; one request's Server-Timing app;dur=75.6 … kpis;dur=1.4, panels;dur=0.1, feed;dur=59.0, render;dur=7.1. The same probe at 074a969f (where the test was added), 362cfdff, fddaa1fa, 97730091 and 79e39296 gives the same spread (owner p95 130–600 ms), so nothing regressed: the bound is end to end, the feed takes most of the time, and the KPI row and pinned panels the test is about take about 1.5 ms.
Done:
[ ] the pull names the cause (from the flake list: a tight wall-clock bound, or a budget measured end to end instead of in the code it guards)
[ ] bun test --rerun-each 30 test/home-cache.test.ts passes, and also passes with other test files running beside it. The pull says how it was checked
[ ] the test still fails when the KPI or panel cache is broken (for example, when it misses on every warm request); what it promises about the product is unchanged. A time budget may only move to what the code itself spends (the kpis and panels phases, or the statement count of a warm hit), never just get wider; docs/CACHING.md §Budgets says what the / test bounds
Files: test/home-cache.test.ts, docs/CACHING.md §Budgets; the source only if the flake is a real slowdown there (then the pull says so).
Intent:
warm home holds p95 under the stated budgetin test/home-cache.test.ts goes red in pulls that never touch home (pull #5079, a two-line docs/DESIGN.md change, check run 01m5hc81ck9rrt6e7qhs33wtdd: p95 698 ms against 200 ms, and still red when the shard gate re-ran the file alone), so good work is parked.What was measured (bench fixture, sandboxed crew machine,
timeRoutewarm, 20 requests): owner home p50 ≈ 95 ms, p95 ≈ 200 ms; one request's Server-Timingapp;dur=75.6 … kpis;dur=1.4, panels;dur=0.1, feed;dur=59.0, render;dur=7.1. The same probe at 074a969f (where the test was added), 362cfdff, fddaa1fa, 97730091 and 79e39296 gives the same spread (owner p95 130–600 ms), so nothing regressed: the bound is end to end, the feed takes most of the time, and the KPI row and pinned panels the test is about take about 1.5 ms.Done:
bun test --rerun-each 30 test/home-cache.test.tspasses, and also passes with other test files running beside it. The pull says how it was checkedkpisandpanelsphases, or the statement count of a warm hit), never just get wider; docs/CACHING.md §Budgets says what the/test boundsFiles: test/home-cache.test.ts, docs/CACHING.md §Budgets; the source only if the flake is a real slowdown there (then the pull says so).