CL-6451: one room participant, one live run — command paths reuse resident agents; stale turns fail

[TheGreatAxios]

Root cause

Two independent defects, one wedge:

1. A mention could mint a sibling run. A participant's mention handle derives from its definition's display name (Myra -> @myra), while the workflow-command registrar names commands after the definition's wire name (assistant). dispatchWorkbenchCommand's known-handle guard compared the handle against the command name, so @assistant sailed past it into dispatchAtCommand -> startWorkflowCommand -> launchAndJoinAgent, minting run B for an agent the room already held as run A. Later messages fanned out to both, run A never parked, and the dispatch died at the supervisor's waitForRunTerminalOrPark 300s backstop. The same class hit /name re-invocations and any handle drift (dedupe suffixes, slugging).
2. A failed dispatch left the turn row running forever. dispatchTurn closes its row only when sendMail throws; a dispatch that fails later in the workflow-host supervisor (the backstop) emits no message.run.ended, so nothing ever closed the row — stuck typing indicator, no failed-turn strip.

Fix

• dispatchWorkbenchCommand now resolves an @name command against the definitions the room's agents were launched from (findResidentAgentForDefinition, comparing by definition asset so re-deployed rows still match). A resident match returns a routing decision instead of a command dispatch: the message posts normally and sendWorkbenchMessage forces the resident participant into the recipient set, so the turn rides the ordinary pipeline — turn row, one-in-flight queueing, reply correlation — into the run the room already has.
• startWorkflowCommand applies the same residency rule to slash/command dispatch: a resident definition gets the args delivered as mail to the existing run, never a second launch.
• The explicit invite affordance deliberately still always launches — a second instance of one definition per room stays possible (that is what handle de-dup echo/echo-2 exists for, and what the CL-6329 live proof pins), guarded by a new test.
• Both AgentTurnStore implementations now fail any row still running past AGENT_TURN_STALE_MS (per-occurrence timeout + grace) on their next read or write — an interrupted dispatch surfaces as a visible failed turn instead of zombieing, and a later reply can never be attributed to a dead row.

Reproduced at scale (longevity campaign, live forensics)

The longevity campaign wedged on exactly this defect while the fix was being built. Read-only forensics from its scratch DB (longevity_scratch), captured mid-wedge:

• One room, 17 agent participants, 17 minted runs — 4 intended agents plus 13 duplicate daily-heartbeat participants (daily-heartbeat through daily-heartbeat-13), one minted per scheduled command invocation (18:00, 18:01, 18:04, 18:27, 18:47, 19:00, 19:06, 19:44, 20:00, 20:07, 20:32, 21:00) — the startWorkflowCommand mint point this PR removes.
• 909 running turn rows, 849 older than 10 minutes, 0 ever marked failed — the zombie-row defect this PR's stale-turn expiry closes. All 905 wedged rows sit on the one working agent (sales-analyst).
• Progressive collapse in lockstep with duplication: sales-analyst completed 77 turns in hour one, then 7, then 2, then 0 after ~20:09, while ~300 new turns/hour kept opening. Hub and sidecar both saturated (~90% CPU in lock waits + replay churn) with zero inference starts.
• The duplicates' opening mails also bypass the turn projection entirely (no turn rows at all for the 13 siblings). With this PR the invocation reaches the resident run instead of minting; the @name form rides the full turn pipeline (turn row, queueing, reply correlation).

Verification

• Red/green at the routes/service/store layers: mention-of-resident routes into the existing run and opens turn__0/turn__1 on the same occurrence sequence; command reuse by id and by asset; stale-turn expiry in both stores (drizzle suite run against a scratch Postgres).
• Live proof on one real stack (scripts/e2e/cl-6451-single-run-proof.ts, scratch DB + real Ollama): invite assistant (handle myra), send @assistant — no command interception, one agent participant, reply as turn__0 of run A; a second @assistant message answers as turn__1 of the same run and recalls a fact stated in turn 1, proving the bootstrap exchange lives in the same stepId-keyed durable history (CL-6453's symptom).
• bun run check green.

Fixes CL-6451
Fixes CL-6453