test: enforce simulation and render budgets

BUILD_PLAN.md

@@ -47,8 +47,8 @@ art; every shipped dependency, route, font, and asset has recorded provenance.
 
 ## M1 — California roads and passive Model X traffic
 
-Status: **implemented and rendered on desktop and mobile; performance-budget
-instrumentation remains**.
+Status: **implemented and rendered on desktop and mobile; repeatable California
+and office browser performance-budget instrumentation is implemented**.
 
 - Coarse California board plus detailed Bay Area and SoCal boards.
 - Serializable US-101 and I-5/I-580/I-80 route graphs.
@@ -137,6 +137,17 @@ validation, and no regression to ambient live ADS-B rendering.
 
 ## Performance gates
 
+Run `npm run build && npm run performance`. The harness serves the production
+bundle locally, uses system Chrome (real GPU when available, SwiftShader as the
+headless fallback), warms each scene, and writes `/tmp/tera-performance-budget.json`.
+Budgets live in `scripts/performance-budgets.json`; a threshold, console error,
+insufficient sample, or anonymous private endpoint request makes the command
+non-zero. Use `--output`, `--budgets`, `--warmup-ms`, `--sample-ms`, or
+`--software` for explicit CI/diagnostic runs. The named matrix is California and
+Office at 1440×900 desktop and 390×844 mobile viewports. Frame comparisons allow
+only 0.1 ms for Chrome's rAF timestamp quantization; the declared and reported
+budgets are not raised.
+
 - Named benchmark scenes: p95 frame at or below 16.7 ms desktop and 33.3 ms on
   the selected supported mobile tier.
 - Hard budgets per scale for resident cells, triangles, draw calls, dynamic