fix(sweep): a test that rotted behind a wrong excuse, and a box spec off by 2x

[catomean]

Follow-ups from sweeping the earlier findings. Two of the four I carried forward did not survive checking and are dropped: /api/control/dispatch is a real route (so the docs referencing it are correct), and scripts/test/push-notifications.ts exists. Verifying first is why this PR is smaller than the list that prompted it.

1. A test nobody ran, for a reason that was never true

scripts/test/push-notifications.ts read scripts/agent-hook-bridge.sh — deleted 2026-06-11 in 956ccf64 ("delete the bash daemon and bridge files"). It threw ENOENT and the suite exited 1.

Its entry in test-unit.ts SKIP said "needs push/web-push env — run manually". That was never the reason: every check in the file is a static file read and it needs no environment at all.

A wrong skip reason is worse than no skip — it answers the question that would have found the bug. The red looked accounted for, so nobody looked again for three months.

Retargeted, not deleted. The behaviour moved rather than went away: a closing run reaches the operator via notify-close.ts → pushToUser → push-fanout. And scripts/test/notify-close.ts covers that module without mentioning push at all — so "an agent finished, tell the operator" was asserted by nobody. Deleting the check would have quietly ratified that hole.

Removed from SKIP, so it now runs in test:unit (132/132) and therefore in CI.

Mutation-proven: renaming the pushToUser call → exit 1, notify-close must call pushToUser. Restored → exit 0.

2. The box is half the machine the docs describe

Measured 2026-09-04: nproc 4, free -m 7746 MiB, df -h / 75 G, swap 4095 MiB. That is a CX33.

Three docs said CX43, 8 vCPU / 16 GB, 40 GB disk — every figure but the swap wrong, in the direction that matters: planning against 16 GB on an 8 GB box is how you meet the OOM killer.

src/app/api/system/hetzner/route.ts has said "the box is a cx33" all along — code and docs disagreeing, with the code right.

Note the shape of the drift: CPU and RAM went down while disk went up (40→75), which is what a disk-inclusive rescale to a smaller tier looks like — plausibly to escape the disk-full problem the old caveat describes. The June runbook entry is kept as history and marked superseded rather than rewritten.

The docs now point at /api/system/hetzner instead of restating numbers. A hardcoded spec is a second source of truth that rots silently — exactly what happened between June and September.

3. The composer status, finished

#479 gave this line its own row below sm but kept sm:truncate, leaving it at 57% on a 1440 viewport — "Autopilot on: queue is empty, so FleetCr…". Same defect as the phone's 14%, just less obvious, and there is no title here so the remainder is unreachable at every width. It now wraps everywhere: costs one toolbar line when the sentence is long, where truncating cost the half of the sentence that says what will happen.

Verification

pnpm run verify — exit 0 (repo moved to pnpm in #478; installed with --frozen-lockfile first).

🤖 Generated with Claude Code

https://claude.ai/code/session_01UvjGNAS9CMfEGNW26tUR4P