feat: add core durable signals and waits (Phase 1)

[hacetin]

Summary

Implements Phase 1 only of #79: authenticated, typed durable signals and signal waits. This partially addresses the issue; it does not close the full webhook/connector roadmap.

• Add validated version-1 signal envelopes, bounded durable FIFO buffering, deduplication, and attributed untrusted-data delivery.
• Add wait_for_signal with indefinite or timed waits, interruption/re-arm preserving the original deadline, and wait-ID-scoped Stop/cancellation.
• Expose start-aware raiseSignal and redacted state inspection through SDK/direct/web clients, authenticated HTTP/RPC, MCP, and tuner tools. Existing event APIs become typed-signal compatibility wrappers rather than raw queue writes.
• Keep the TUI and portal aligned on pending signal names, deadlines, interruption, and lifecycle Activity/sequence entries; update the proposal, canonical docs, builder guidance, and sample documentation.

Replay and rollout safety

Merged main at 7cac3109, retained its non-signal orchestration 1.0.79 as a frozen handler, and introduced signals under 1.0.80 with pilotswarm.signals.v1 capability-routed turn activities. Signal writes and inspection reject 1.0.79 and older decoders. Existing main executions retain their original handler until a continue-as-new upgrade boundary; old activity descriptors and tool declarations remain unchanged.

The frozen 1.0.79 files match main with only their current-version constant pinned. The existing 1.0.78 frozen copy is unchanged and no longer appears as an addition relative to current main.

Draft-lab compatibility: the earlier unmerged draft used 1.0.79 for signals before main independently assigned that number to different behavior. Those draft-test histories are not main's 1.0.79 histories. Use a fresh isolated test database when moving those experiments to this corrected build; ordinary main histories remain supported.

Review corrections

• Resolve the orchestration-version collision and enforce the new 1.0.80 signal-support floor.
• Examine queued input before committing an overdue signal timeout, including after a delayed activity; enforce a named agent's required first tool on a signal-first turn.
• Preserve saved user input and budget-retry timers during signal bursts, and let a matching wake satisfy the saved waiter without violating oldest-matching FIFO order.
• Resume configured interval and wall-clock cron schedules after stopping a signal wait, without replacing another active budget-retry timer.
• Expose the existing Stop control for valid parked signal waits in the portal/TUI, and preserve authoritative wait metadata across CMS-only list refreshes while still clearing ended waits on current rich snapshots.
• Bound generated reasons so eight maximum-length valid signal names remain inspectable.

Validation

• Full workspace build passed after merging main.
• 1,051 SDK unit tests and 42 API tests passed.
• All app TUI, portal, shared UI, and MCP regression suites passed.
• 209 focused/native/adjacent orchestration tests passed across 10 files, including seven native Duroxide cases.
• Native regressions cover replay of main's 1.0.79, rejection of signals on its old decoder, continue-as-new to 1.0.80, pre-deadline signal acceptance across a delayed waiting activity, matching budget-interrupted wakes, Stop/cron continuation, and the existing worker-replacement/buffering coverage. Native fixtures use real Duroxide/SQLite with stub model activities; they are not credentialed PostgreSQL/Copilot tests.

Pending before readiness: the full credentialed ./scripts/run-tests.sh gate is blocked because the worktree has no .env with DATABASE_URL and GITHUB_TOKEN. Secret configuration was not created or modified; native fixtures do not replace live PostgreSQL/Copilot coverage.

The manual section below is preserved verbatim from the user's tests of the earlier PR head fe675a87; it does not claim that the corrected 1.0.80 build has been manually re-tested.

Manual Local Testing

HAPPY PATH:

1. Start a session send the prompt to both sessions (main and this branch)
2. Wait a few seconds and send the same payload to both

PAYLOAD='{
  "signalId":"manual-build-1",
  "data":{"buildId":"B-42","status":"succeeded"},
  "wake":false
}'

# Main: expect HTTP 404 because this route does not exist.
curl -sS -o /dev/null -w 'main HTTP %{http_code}\n' \
  -X POST "$MAIN/api/v1/sessions/$MAIN_ID/signals/build-finished" \
  -H 'Content-Type: application/json' -d "$PAYLOAD"
main HTTP 404

# PR: expect a queued receipt.
curl -fsS \
  -X POST "$PR/api/v1/sessions/$PR_ID/signals/build-finished" \
  -H 'Content-Type: application/json' -d "$PAYLOAD" | jq
{
  "ok": true,
  "result": {
    "signalId": "manual-build-1",
    "name": "build-finished",
    "raisedAt": "2026-09-18T14:21:12.863Z",
    "status": "queued"
  }
}

Before:

After:

Also tested:

• Indefinite wait/resume and duplicate suppression passed.
• Pre-arrival buffering passed; FIFO consumption order remains untested.
• User interruption preserved the deadline; timeout fired once.
• wake:true woke an idle session; Stop cancelled waiting while preserving buffered signals.
• A parked wait survived an actual process restart.
• The 32-signal cap, audited oldest-first overflow, invalid-name rejection, and oversized-payload rejection passed.

Not included

wait_for_any, public webhook endpoints, GitHub/Azure DevOps connectors, trusted event bindings/session templates, coalescing, and webhook receipt/dead-letter operations remain later phases of #79.