refactor(commands): extract command poll + dispatch + ack into commands/drain.ts

[itskai-dev]

Summary

§11e step 1 of COMMAND_PLUGIN_SKILL_ARCHITECTURE.md v1.0 (reflectt-cloud merged at da6cdba3). Behavior-zero relocation of the inline command runtime that lived in cloud.ts into a dedicated src/commands/drain.ts.

This is step 1 only. Per @kai (msg-1777492135899): "After that, reassess before the empty registry PR." Step 2 (empty registry contract + relocate the two existing verbs into per-verb handler files) lands as a separate PR.

What moves

From src/cloud.ts → src/commands/drain.ts, verbatim:

• pollAndProcessCommands() — heartbeat-tick poll, 10s active / 60s idle
• handleCommand(cmd) — inline if/else dispatch by cmd.type
• handleContextSync(cmd) — fetches local context, posts to cloud sync
• handleRunApprove(cmd) — calls local approval-queue, acks
• COMMAND_POLL_ACTIVE_MS / COMMAND_POLL_IDLE_MS constants
• commandPollErrors / lastCommandPollAt counters
• PendingCommand interface

cloud.ts exports state, config, isIdle, cloudGet, cloudPost so drain.ts can import them. ES live-binding semantics preserve behavior — cloud.ts still owns the lets and reassigns config at line 368. The two existing pollAndProcessCommands() call sites in cloud.ts (heartbeat tick + immediate kick) are unchanged.

Scope locks (kai msg-1777492135899 / link msg-1777492193718)

• behavior-zero
• same poll cadence
• same ack contract
• same inline dispatch semantics
• no registry creep until step 2
• no auth/intake churn
• no new verbs

Gates

• npx tsc --noEmit — clean
• npm test — 242/242 files pass, 2608/2608 tests pass (1 pre-existing skip), 31s
• npm run build — clean
• canonical staging command-queue smoke (poll, ack-complete, ack-fail) — recommend rolling rn-34faba44-wlgkeq to the new image and walking a context_sync command before merge

Test plan

• npm test green in CI
• npm run build green in CI
• After merge: docker-publish rolls new image; canonical host rn-34faba44-wlgkeq updated; observe 📬 Command polling: 10s active → 60s idle startup log + at least one ☁️ [Commands] Processing context_sync line in production logs

Reviewer hooks

• @link — runtime/drain seam: is commands/drain.ts the right module shape? Are the cloud.ts re-exports of state/config/cloudGet/cloudPost minimal enough?
• @kai — confirm step 1 alone (reassess before step 2 per your msg-1777492135899)

🤖 Generated with Claude Code

[itskai-dev]

Closeout — §11e step 1 merged on 2/3 proofs

Per kai (msg-1777493903222) + link (msg-1777493906404):

Proofs captured on canonical staging (rn-34faba44-wlgkeq machine 568354e7ad5348 running ghcr.io/reflectt/reflectt-node:6aca844):

1. ✅ Clean boot — agent suite (compass/main/orbit/watchdog) starts normally
2. ✅ No regression — proactive ContextSync, presence/events/chat all healthy
3. ❌ One real command through the relocated drain — upstream blocker, not a drain concern

Named seam (carries to step 2 reassess, not into this PR):

The cloud API has no HTTP enqueue endpoint for commands. apps/api/src/routes/registry.ts registers only GET /api/hosts/:hostId/commands (poll) and POST .../ack — there is no POST /api/hosts/:hostId/commands. The in-memory commandStore (apps/api/src/domains/command/queue.ts) has exactly one in-process producer: spend-alert.ts calling pushCommand on threshold breach.

So canonical can prove clean boot + no regression, but not a naturally-queued command through the relocated drain path — because no reachable cloud enqueue path exists yet.

Next reassess (per kai): decide whether "no reachable cloud enqueue path" belongs inside the next node-runtime step (§11e step 2 = empty registry + per-verb handlers), or is a separate cloud/runtime seam that stays split out.

Until that call: hold per feedback_stop_and_hold_after_lane_close.