fix: harden queue lifecycle and publish gate

- Preserve phase results on partial retry and keep interrupted phase
  context after restart.
- Avoid webhook bookkeeping crashes when retention deletes stale jobs.
- Add deeper unit, integration, and e2e coverage around queue seams.
- Require verify job to pass before publish runs.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

docs/README.md

@@ -0,0 +1,39 @@
+# jobqueue docs
+
+Detailed architecture and runtime docs for `jobqueue`.
+
+## Doc map
+
+| File | Purpose |
+| --- | --- |
+| [`architecture.md`](./architecture.md) | Static architecture, module boundaries, data model, event model, state model |
+| [`runtime-lifecycle.md`](./runtime-lifecycle.md) | Step-by-step runtime behavior from startup through shutdown |
+| [`integration-findings.md`](./integration-findings.md) | Multi-agent scan results, verified bugs, fixes, and remaining behavioral notes |
+| [`jobqueue.c4`](./jobqueue.c4) | LikeC4 source for landscape, container, component, and runtime views |
+
+## Mental model
+
+`jobqueue` is a **single-process orchestrator** around a SQLite-backed job table.
+
+1. Consumer code creates a `JobQueue`.
+2. Jobs are persisted immediately in SQLite.
+3. A pump claims runnable jobs and hands them to a concurrency-limited worker pool.
+4. `PhaseRunner` executes configured phases and reports progress back through `JobQueue`.
+5. `JobQueue` persists each state transition, emits typed events, formats SSE payloads, and optionally sends webhooks.
+6. A retention scheduler can mark old jobs as `stale` and later delete them.
+
+## Rendering LikeC4 views
+
+The repository stores LikeC4 source in [`jobqueue.c4`](./jobqueue.c4).
+
+```bash
+npx likec4 start docs/jobqueue.c4
+```
+
+Recommended views in the file:
+
+- `index` - system landscape
+- `library` - container view of `jobqueue`
+- `runtime` - internal runtime/component view
+- `enqueue-to-complete` - dynamic happy path
+- `retry-flow` - dynamic retry path