2026-05-29 — Simulation gateway (Slack-free verification)

Decision: Extract a Transport port and invert createSlackApp → createGateway(agent, transport). Production binds bolt; simulation binds an in-memory SimTransport. All gate/command/approval logic runs unchanged in both — consistency by construction (the full existing suite stayed green across the refactor: 594 tests).

Why not fake the Slack wire protocol: adapter.ts hardcoded new App(), so it needed a refactor regardless; faking bolt's receiver + action envelopes is brittle across bolt bumps and is itself untested surface. Inverting the edge is mechanical and lower-risk. Bolt's App satisfies the Transport port structurally, so the production path is untouched.

Stub agent seam: StubAgent drives turns by calling the REAL exported MCP handlers (slackHandlers, brokerHandlers) through a narrow __sessionCtx seam on the gateway handle, so approval-gate and the connect-broker execute for real without Claude.

Real-behavior facts the sim pinned down:

• The soul regex fallback (extract.ts) only fills approvers; without the LLM, manager and channel allowlists come back empty. The sim injects structured SoulData via the existing production setSoulData() accessor (no new production seam).
• The ApprovalGate authorizes only the catchall approver (U0APP); the manager is NOT implicitly an approver — a manager click leaves the card pending.
• Channel (non-DM) inbound is dropped for lack of an @mention before the channel-mode gate runs; the sim injects the bot mention to engage the thread (mirrors a real @mention).

Drop assertions: the runner wraps the existing metric.slackDropsTotal.inc counter to capture {reason} labels, restoring it on dispose.

Deferred: file attachments (fetch-based, off the WebClient surface), web UI, real-agent CI runs, full owner-approval borrow transcript (needs a connect step + the CDP login-host seam; the borrow-grant scenario currently proves the broker tool is reachable and returns a needs-connect hint).