chore(gastown): promote gastown-staging to main

[jrf0110]

Summary

Batch promote of gastown-staging to main. Four commits since the last promotion, each addressing a distinct gastown issue:

#	Commit	PR	Type
1	21a14d04	(no PR — direct push)	dev-only fix
2	c532f40e	#2999	bug fix
3	db584066	#3047	bug fix
4	47e95a8b	#3055	observability + race fix

Constituent changes

• fix(gastown): point dev GIT_TOKEN_SERVICE binding at git-token-service-dev (21a14d04, no PR — direct push)
  Local-dev binding fix. services/git-token-service/wrangler.jsonc overrides its worker name to git-token-service-dev in env.dev, but gastown's env.dev.services binding still referenced the base git-token-service name. Wrangler's local dev registry does exact-name matching, so the binding showed [not connected] whenever both workers ran side by side. Production binding (top-level services) is untouched; only the env.dev block was changed.

• fix(gastown): push new model onto resumed mayor session on hot-swap (#2999)
  Regression where changing the mayor's model in town settings persisted the config but the running mayor kept using the previous model — users had to run /model manually. Root cause: a prior fix (9785570b9) skipped the session.prompt(...) call on resumed sessions to avoid a duplicate startup turn, but that call was also responsible for pushing the new model field onto the session. Fix sends a session.prompt with noReply: true and the new model on resume, so the model updates without injecting a synthetic user turn.

• fix(gastown): stop reconciler log spam from orphaned bead_cancelled events (#3047)
  Production logs were filling with reconciler: applyEvent failed … Bead <id> not found repeating every alarm tick forever. Two cooperating bugs: (1) deleteBead cascaded cleanup to satellite tables but not the town_events reconciler queue, leaving orphan events; (2) the drain loop in Town.do.ts intentionally never marked failed events as processed, expecting retries to eventually succeed — but for a deleted bead they never can. Fix cleans up town_events in deleteBead/deleteBeads, pre-checks bead existence in applyEvent, and classifies Bead/Agent … not found errors as terminal in the drain loop.

• feat(gastown-container): add crash visibility + per-agent start mutex (#3055)
  Investigation hooks for repeated container restarts seen on a specific town (~1.5–2 min cadence). Adds an unhandledRejection listener with full stack logging (no process.exit), periodic RSS memory logging via the heartbeat path, and a per-agentId mutex in startAgent that fixes a real concurrency race exposed by duplicate /agents/start log lines arriving in the same millisecond for the same mayor.

Verification

• Each constituent PR was reviewed and merged into gastown-staging independently — see PR links above for per-change verification details.
• 21a14d04 was manually verified locally (wrangler dev --env dev for both workers, binding now connects).

Reviewer notes

• All four changes are scoped to services/gastown (container + DO/worker code) and the env.dev block of services/gastown/wrangler.jsonc. No cross-service changes.
• Production worker bindings are unchanged.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (9 files)

• services/gastown/wrangler.jsonc
• services/gastown/container/src/main.ts
• services/gastown/container/src/process-manager.ts
• services/gastown/container/src/process-manager.test.ts
• services/gastown/container/vitest.config.ts
• services/gastown/src/dos/Town.do.ts
• services/gastown/src/dos/town/beads.ts
• services/gastown/src/dos/town/reconciler.ts
• services/gastown/test/integration/event-cleanup.test.ts

---

_{Reviewed by gpt-5.5-20260423 · 550,886 tokens}