Remaining fleet proofs against schema 15

2026-09-27 continuation of VUH-1381. Managed Swarm owner: b62a68f, schema 15. No Rivals runtime or lead changed.

• Prompt/wait: started the reserved Codex scratch agent in w2:p2K; the routed CLI prompt/wait returned done and the pane answered VUH1381_PC_PROMPT_WAIT_OK.
• Desktop session: the existing desk-agent task was Ready, with only our previous harmless job queued. Started that existing task once. The job output reports its own PowerShell PID 157536 in session 1 and console Active. No Herdr server started or replaced.
• Persisted watch: armed 33f95216-0505-4f2e-b487-dbc94cdb9dc7 at 18:00:49.874Z, then asked the lead for a restart. The same record appears before and after the lead's 18:04:05Z restart. Only then released the marker. The scratch agent returned VUH1381_WATCH_SURVIVED. The lead confirmed delivery into the original conversation at 18:04:34Z, event seat-0583a701-b5de-4262-a103-f4b25169c4e5 (receipt). The original watch was subsequently absent, consistent with one-shot delivery. No replacement watch was armed.
• Remote hire: first exposed missing promptAgent forwarding in the combined fleet runner. Added forwarding and regression coverage; lead loaded it in the same restart. The next Codex hire failed agent_not_running, so Codex remote hire remains unverified. A Claude hire succeeded on pc/term_65c7aca90366f4a, pane w2:p2W, delivered its brief, and answered VUH1381_REMOTE_HIRE_OK. Closed only that newly hired canary after preserving the result. Failed hires clean up their newly created panes automatically; the borrowed scratch pane remains open.
• Schema-15 relay: new enrolled PC actor 766c124c-9211-4f42-99c3-6300f0ef8ea9 authenticated against the same scope. Old 0981253 correctly refused the changed schema. Staged b62a68f separately and authenticated. The PC acknowledged nonce 7d065c95; the Mac acknowledged reply ca92a6e7 on attempt 1 (events). Adapter staging reused the existing installed dependency tree. The new package also declares MCP client 2.0.0; a full clean managed-worker installation is not claimed by this adapter test. Capability files stayed private and are excluded from evidence.

The named fleet retains its explicit canary directory grant. The desktop bridge is running. The scratch agent is idle after completion. The relay remains service-supervised. No lead migration or crash test was executed; that is the acceptance criterion that remained at the time, with a runbook prepared earlier. The coordinator handoff procedure was later retired by ADR 0213.

Validation

All 12 focused fleet tests pass, including first-prompt routing. Full pnpm check reached tests after formatting, lint, dead-code, docs, infra and typecheck checks. Vitest finished with 336 files passing, 2815 tests passing, one skipped, and one failed release test. The release test independently fails because the installed Swarm dependency tree cannot resolve the newly declared @modelcontextprotocol/client dependency; diagnostic. The repair integration owner must install the updated graph and recheck it; this worker does not change vendor provenance or the lockfile.