[jackulau]

The site has not deployed since #32 merged. The Site workflow stops at its Perf budget step: /docs/api/read/ is 30.3KB brotli against a 30.0KB budget. The page is the part of the generated API reference that gained the weight edit and checkpoint fields.

What changes

• Budget. /docs/api/read/ goes from 30KB to 33KB in apps/site/scripts/perf-budget.mjs, about 10% over what the page measures, the same rule the other generated reference pages follow.
• CI. The budget ran only in the Site workflow on main, so feat: weight edits read the same on every surface, and experiments record their checkpoint #32 passed every check and then stopped the deploy. CI now runs npm run perf:budget after the site build on the Linux Node 24 leg, the same OS and Node as the Site workflow. The step took 22 seconds in the Site workflow.
• Docs. docs/perf.md says where the budget runs.

Two commits, one per change.

Checks

• The old script on a build of main's content: FAIL /docs/api/read/: HTML is 30.3KB brotli, budget 30.0KB. The new script on the same build: perf-budget: ok, closest page /docs/agents/ at 25.1KB of 26.0KB.
• actionlint -shellcheck= .github/workflows/*.yml: clean. npm run lint: clean. Docs tests: 51 of 51.

Not checked locally

• The new CI step itself; this pull request's own run is its first.
• The site was not rebuilt here. The budget ran against a build made from fix: hook tests pass inside an agent, and the Python weights warning is once per process #33's head, which has the same tree as main.

[jackulau]

Correction to the description: "the site has not deployed since #32" overstates it. The Site workflow failed at Perf budget on the #32, #33 and #34 merge commits, and it is green again on this merge (154558e6). Its Deploy step is skipped on every run, before #32 as well, because the CLOUDFLARE_API_TOKEN secret is not set, so https://coding.sh still serves the last deploy made by hand.