A deploy records the moment it started serving, and the studio says how long that took #30

Already up to date (the commit was pushed earlier before the budget cutoff). Work is landed.

A deploy run now gains servingAt, stamped once at the swap-in — the same instant the runtime writes serving <url> — never at the moment its tests finish beside it.

Done:

• Migration 22 adds run.serving_at (nullable, set once); core/runs.ts exposes it on Run and adds markServing.
• apps/runtime.ts stamps it at the swap-in and the log line now reads serving <url> — 3.4s after the deploy started.
• The studio page's app head and a pull's preview block say live 3.4s after the push, starting before it has served, and nothing (no zero) if it never did.
• A deploy whose tests then fail keeps its failed run and red check exactly as before; only the words changed, to serving · tests failed instead of anything readable as "not up".
• servingAt rides along on the run wherever the full Run/EnvStatus objects already serialize to JSON (/runs, /runs/:id, the app's environments) — no api.ts change was needed since nothing there hand-picks fields.
• docs/APPS.md documents the new field and its relationship to startedAt/finishedAt.
• Two new assertions in test/apps.test.ts: the existing "slow test step" deploy now also checks servingAt is set and earlier than finishedAt; a new test drives a deploy that never answers and checks servingAt stays null (and closes its pull so it doesn't leak into the later restart-count test).

Verified: bun x tsc --noEmit is clean; the full forge.json check suite ran 182 tests, 177 passed. The 5 failures are all the same pre-existing sandbox limitation — local git pushes over HTTP failing with 128/auth required — in tests untouched by this change (apps-adversarial.test.ts, cross-repo-pull.test.ts, web.test.ts, pulls.test.ts, and the same known failure inside apps.test.ts itself); this reproduces identically on main before this change.

Worth a look: the exact wording choices — "live 3.4s after the push" and "serving · tests failed" — are judgment calls where the issue gave an example rather than a literal string; the reference instant used is the deploy run's own startedAt, per the issue's risk note, not the push event itself.

Budget reached at $4.26; this is where it stopped.

Closes #26

Pushed. Merged main into this branch, resolving both conflicts:

• src/core/schema.ts: main added migration 22 (the decision ledger); mine added a migration for run.serving_at. Kept main's 22 as-is and renumbered mine to migration 23, so both land in sequence.
• test/apps.test.ts: same test had grown two independent tails — my servingAt/finishedAt ordering assertions, and main's added pull-merge cleanup (so a later test only sees the preview it opens itself). Kept both, in that order.

bun x tsc --noEmit passes clean, and the full forge.json check suite runs with only pre-existing failures (git-push-over-HTTP fails with "auth required" in this sandbox) — verified identical on origin/main alone, so unrelated to this merge.