[DECISION] /turn-stream is not retirable — the broadcast half is, the turn lock is not

[drewstone]

Answering "can /turn-stream + TurnStreamDO be retired now that both platform lanes fan out to the session gateway?"

Verdict: NO. Not fully, and not yet even partly.

/turn-stream carries four responsibilities. The session gateway replaces exactly one.

Responsibility	Gateway replacement?
Live turn-event broadcast + reconnect replay	Yes — for sandbox turns, once #4114 reaches production
Single-flight turn lock (createDurableTurnLock)	No — the gateway has no mutual exclusion
Durable turn-event store (createDurableObjectTurnEventStore)	Partly — durable chat rows (#250) serve history, but the store also backs the app-side replay endpoint
Product extension seams (handleProductRequest / shouldDeferLockRelease / productSyncEvents)	No — product domain, no analogue

The lock has no substitute

The only thing named lock in the gateway is SubscriptionMutex (apps/orchestrator/src/session-gateway/sse-bridge.ts:42) — private locks = new Map<string, Promise<void>>(), an in-process guard against duplicate SSE subscriptions on one instance. It goes silent across instances, the same defect as the platform's in-process workflow-run-events.ts emitter. It is not a distributed turn lock and was never meant to be.

Sizing the retirable slice

gtm-agent is the sole consumer. src/lib/.server/chat/session-broadcast.ts imports 11 symbols from /turn-stream; only 3 are broadcast (broadcastTurnStreamEvent, broadcastWorkspaceActivity, broadcastThreadCreated). The other 8 are the lock and its types. SessionStreamDO (758 lines) additionally hosts the Vault mutation/head endpoints, the interaction-answer rendezvous, and the vault-turn persistence queue.

Retiring the broadcast half removes 3 of 11 imports from one file. The Durable Object stays.

Preconditions before even the broadcast half can go

1. agent-dev-container #4114 reaches production. Merged to develop (a0adc3580); git merge-base --is-ancestor a0adc3580 origin/main -> NO. Production sandbox.tangle.tools still delivers 0 turn events on the run/stream (detach) lane. The main <- develop release PR is adc #4108, open.
2. gtm adopts agent-app >= 0.43.62. Currently pinned ^0.43.60 — before feat(chat-routes): persist assistant parts incrementally as a turn streams #250, so it has no incremental persistence and no durable history for late viewers.
3. The 2.000x lane-switch defect is located. Measured on production: feeding session-bus events to the shell's producer wrote a durable row exactly 2.000x the true transcript. It is not in createSandboxChatProducer — tests/chat-routes/lane-portability.test.ts pins that at 1.000x across eight lane shapes, and removing the reconciliation makes it 3.133x, so the probe is sensitive. The defect is somewhere at the gateway-frame -> producer boundary and is currently unlocated. Switching gtm to the message lane before finding it doubles every stored transcript.

Docs

The measured model is recorded in AGENTS.md § "Live viewing vs history" (#252).