The browser performance runner loads deterministic checkpoints from `test/fixtures/effect-visual-checkpoints.json` and exercises all six reaction styles at normalized lifetimes of 0.10, 0.25, 0.50, 0.75, and 0.95. Direct main-thread rendering with glyph and path caches bypassed is the correctness reference. Cached output may differ only within the anti-aliasing tolerance, with at least 99.5% of pixels inside that tolerance. Structural reductions, style substitution, dropped overlap, shorter lifetimes, or automatic quality tiers are release failures.
Normal-speed Edge gates enforce the 30 FPS cadence, per-style work limits, the four-effect overlap budget, full rendering of eight simultaneous valid effects, bounded scheduler callback rates, reaction publication p95 at or below 2 ms, Gem and completion setup p95 at or below 3 ms, localized inventory patch p95 at or below 8 ms, exact inventory focus/scroll/node retention, Aurora active-write and inactive/hidden zero-work behavior, effect-plus-pan/zoom/pickup/edge-pan operation, and zero owned state after a 100-cycle cleanup stress. The settled heap delta must remain at or below 2 MB after the bounded caches are cleared. The 4x CPU profile is an overload diagnostic: it enforces bounded callbacks, complete rendering, and cleanup, but it does not permit the test or runtime to reduce visuals merely to meet normal-speed timing.
`effects-performance-smoke-test.js` is a documented repository-policy guard. Its source assertions protect the no-auto-degradation contract, bounded glyph/path caches, shared animation templates, localized inventory patches, and cleanup/lifecycle mechanisms that do not yet have a smaller public module seam. The browser runner always writes its report to `test-results/browser-performance-benchmark.json` unless `BEND_FIELD_BENCHMARK_OUTPUT` supplies another path, including when a gate fails after measurement.
