Pacing decides from the provider's measured percent left, and concurrency goes back to its ceiling #9002

I merged main into the branch and pushed it. The merge had two conflicts, docs/CREW.md and src/crew/week_read.ts, and both are resolved.

• docs/CREW.md: I kept both new sections, "The weekly pace reads the provider's measure (#8960)" and "What the pace limits (#8962)".
• src/crew/week_read.ts: weekFor now passes floorSlots: paceFloor(app.db) from main into the decision. It keeps this branch's non-null week, so I dropped main's week === null check.

bun x tsc --noEmit and bun scripts/gate.ts were clean. The seven test/week*.test.ts files pass (65 tests). I did not run the full suite; the forge's check does that on the push.

Continues #9002

I merged main into the branch and pushed it. The merge had two conflicts, docs/CREW.md and src/crew/week_read.ts, and both are resolved.

• docs/CREW.md: I kept both new sections, "The weekly pace reads the provider's measure (#8960)" and "What the pace limits (#8962)".
• src/crew/week_read.ts: weekFor now passes floorSlots: paceFloor(app.db) from main into the decision. It keeps this branch's non-null week, so I dropped main's week === null check.

bun x tsc --noEmit and bun scripts/gate.ts were clean. The seven test/week*.test.ts files pass (65 tests). I did not run the full suite; the forge's check does that on the push.

The three failing tests in test/pace-floor.test.ts pass now, and I've pushed the fix. I ran that file alone (4 pass, 0 fail) and bun x tsc --noEmit was clean. I did not re-run the full suite; the forge's check does that on the push.

The tests came in with main's #8962 and built their week from dollar spend. Pacing on this branch decides from the provider's measured percent of the week, so those tests had no measure. The pacer fell back to the unknown-week hold, and the review and the person's assignment never started. That is why they timed out at 20 seconds.

The fix is in the test setup only. pacedRepo now records a measure (spentUsd percent used, with the reset 3.5 days ahead, so half the week has gone). I removed the dollar-spend and noteLimit seeding, the recordUsage import and its test-only spend helper. No test expectation changed, and the pacing code is untouched.

I merged main into the branch and pushed; src/crew/week_read.ts was the only conflict. The type check and gate are clean. I ran only the week, pace and measure test files (99 pass, 0 fail), not the full suite. scripts/fast.ts said the diff was too broad to name, so the forge's check runs the rest.

• Imports: I kept this branch's measureFor/getHeadroom and main's capacityCount. standings is gone because the pacer no longer reads dollars.
• Week read: it keeps the provider's measured percent. An account with no measure still gives an unknown week that holds the crew modestly, never "off". I also gave that unknown week the account scope that main's crew-wide pacing (#8989) needs.
• weekFor: main's week === null check is dropped, since the week is never null on this branch.

I did not touch main's weekScope and crewWeekFor. They still handle a null week, which can no longer occur.

Pacing now decides from the provider's measured percent of the week. An unknown week holds the crew at a modest level instead of switching pacing off. crew.concurrency goes back to its ceiling, and pacing is what slows the crew.

Review round, point by point:

1. Beacon tests restored. test/measure.test.ts is back as it was before this pull, covering the 2048-byte script, the origin and size refusals, Do-Not-Track, rate limiting, the daily salt and the daily cap. The pacing-measure tests now live in test/subscription-measure.test.ts.
2. Wiring tests added.

- Stream parser (claude.ts): a fake claude emits a rate_limit_event. It calls onMeasure once for the event that carries a measure and never for the status-only one, and the line does not reach the log. - headroom (usage.ts): it prefers the measure (pct = round(100 - usedPct), resetsAt from the measure). It falls back when there is no measure or the window has reset. - Runner (runner.ts): the new test/runner-measure.test.ts drives two real runs through a fake claude. A run with no grant stores the measure against platform, and a granted run stores it against that subscription.

1. Null branches: I kept them on purpose. weekDecision still returns null when the measured reset is past or more than a week away, which is a real case. So weekFor, crewWeekFor and weekWords can return null, and autonomy.ts, dispatch.ts and the home/terminal lines handle it. What was dead was weekScope's string | null, since weekOf never returns null. I narrowed it to string and corrected the doc comments. The "Never off" wording is gone from weekWords.
2. Shape and frequency: not settled. I could not run Claude Code here, and the repository holds no recorded event, so the fixture lines in the test are written from the documented shape, not recorded. The test says so in a comment. I do not know how often utilization really arrives. If it is rare, the unknown-week hold is the common case, and a person should check a recorded session before this lands. The hold is WEEK.unknownSlots, and the pace sentence says it is unknown.
3. Trade-off to see: test starting state. Every crew test now starts at the unknown-week hold. test/helpers/week-quiet.ts seeds a measured quiet week, and its seven call sites use it so those tests do not have their slots held. The createRun and recordUsage imports in test/week-wiring.test.ts are still used, by its spend helper.
4. Docs. docs/CREW.md now says an admin sees the measure in usage headroom, on the page and over MCP. The terminal reads the same readForgeUsage: the Usage screen and the forge usage block print percent free per window.

Checks: scripts/gate.ts and tsc --noEmit are clean. The new tests plus the week, pace, ledger and crew-held suites pass (about 75 tests). scripts/fast.ts named 1374 files, too broad to run here, so the forge's full check covers the rest.

To try it, open /usage as an admin after a run reports a rate limit. The week's percent free there should match the home pace sentence.