refactor(control-plane): one post-lease agent re-read instead of four copies

[zfy0701]

Summary

Four route files each carried their own refreshMutationAgent — the re-read an agent-scoped mutation takes after it wins the move gate. They move to one http/mutation-agent.ts. Net -39 lines, no new behavior except the drift correction below.

Failure scenario

The copies had already diverged, which is what makes this worth doing rather than tidy.

#1055 corrected three of them from daemonId column equality to placement identity (samePlacementRef) and left the fourth — agents.ts:1343 — on the column. So on main today:

// agents.ts
current.daemonId !== observed.daemonId        // two DIFFERENT member sets both read as null

An agent on member set A, re-placed onto set B while a console edit holds the mutation lease, has daemonId === null on both sides. The columns compare equal, the fence passes, and the edit commits against a placement that has already changed. The three corrected copies would have refused it.

It is a narrow window — it needs a re-placement between the caller's read and the lease — and set→set moves are not common yet, which is exactly why the divergence went unnoticed. The point is that one rule living in four places produced it in a single PR.

Fix

http/mutation-agent.ts holds the rule once:

• placement IDENTITY, not the column — a set placement names no machine, so column equality both misses a re-placement onto another set and reads every set agent as unplaced;
• lastModifiedAt, so an edit that landed between the caller's read and the lease is a conflict.

agents.ts, crons.ts and integrations.ts keep a one-line binding so their call sites are untouched; slack-install.ts's inline copy becomes a call. samePlacementRef is no longer imported by any route file — it now has exactly one consumer.

Test plan

New src/http/mutation-agent.test.ts (unit, no Docker) covers the fence directly: unchanged row admitted and fenced on the observed row's own org; vanished row and edited lastModifiedAt refused; machine placement moved to another daemon refused; set placement moved to a different set refused; unchanged set placement admitted; kind changes refused both directions.

Mutation-checked: swapping samePlacementRef back for current.daemonId !== observed.daemonId fails exactly the set→set case and nothing else.

• pnpm --filter @agentconnect.md/control-plane typecheck — clean
• pnpm --filter @agentconnect.md/control-plane test:unit — 157 files, 1711 tests, all passing
• pnpm --filter @agentconnect.md/control-plane test:int — 89 files, 1363 tests, all passing (this sandbox blocks the Testcontainers image pull at the egress proxy, so the suite ran against a local PostgreSQL 16 cluster through a scratch global-setup with the identical migrate/seed/clone-per-pool contract; that scratch file is not part of the diff)
• pnpm lint, pnpm format:check — clean

Follow-up to #1055.

---

Generated by Claude Code

[agentconnect-md-test]

Approved. I verified the isolated checkout is the trusted synthetic merge whose parents are exactly base b17f981d49d929ca5a44bcbb9b8d66b0907dc31d and head a365557872b2ffdf4a6821b15152dd2c80de0cd6. The extracted helper preserves the existing org-scoped re-read and lastModifiedAt conflict fence while applying full placement identity uniformly, which closes the set-to-set movement gap in agents.ts without changing the corrected callers’ behavior.

Verification completed: control-plane typecheck passed; all 157 control-plane unit-test files and 1711 tests passed; focused ESLint and Prettier checks passed. No blocking findings.

_{sent by review-bot (Codex · gpt-5.6-sol) · open in session}