fix(slack): recover context after a lost first turn

[overlord-letta]

Summary

• recover persisted Slack thread routes whose first user turn never reached the conversation
• use durable user-message history as bootstrap proof, then persist a route marker so healthy threads keep the existing incremental behavior
• cover restart recovery through the real Slack preparation and notification-formatting boundary, including root text, prior human context, and root attachments

Before / after

Before, route existence alone disabled first-turn hydration. If route creation survived but the first delivery did not, a later message such as bump arrived without the original thread context.

After, an unmarked existing Slack channel route checks for a persisted user message. Empty routed conversations re-run first-turn hydration; once a user turn is durable, the route is marked and future turns stay incremental without repeated history probes.

Limits and risk

• scoped to existing Slack channel/thread routes; direct-message and non-Slack routing are unchanged
• backend probe failures preserve the existing incremental behavior rather than speculatively replaying thread history
• legacy routes incur one lightweight user-message probe before receiving the durable marker

Test plan

• bun test src/channels/slack-registry.test.ts src/channels/slack/thread-context.test.ts src/channels/routing.test.ts — 19 passed
• regression test fails against the buggy base when applied without the source fix
• bun run check — all 12 checks passed in an isolated worktree with dependencies installed from the rebased lockfile
• full unit runner: 694 passed; 7 unrelated src/websocket/app-server.test.ts WebSocket error-event tests fail in this environment and were also reproduced from the untouched main checkout

Overlord (agent-c2adbf5c-8419-4211-8cd8-3740db164974)