Feature/observability gateway health (#247)

* feat(observability): revision_count + audit_log query index (migration 045)

Adds tasks.revision_count (the O(1) rework counter — forward-only, existing
rows default 0) and the composite index audit_log(target_id, event_type,
timestamp) that powers the cycle-time and rework reconstruction queries.
Verified the real upgrade/downgrade/upgrade chain on a throwaway pgvector PG.
First task of the 0.10.0 observability dashboards.

* feat(observability): count reworks + attribute qa_fail/pr_fail to the rejector

Every transition into needs_revision increments tasks.revision_count at the
single audit chokepoint (exactly once per bounce, across all paths incl. pr_fail
and ceo_reject), so the rework rate is an O(1) read. A QA or PR-review bounce
also emits a named task.qa_fail / task.pr_fail audit event carrying the
rejector's agent_id, so the per-agent rework scorecard charges the rejection to
the reviewer who made it, not the developer who owns the task.

* feat(observability): cycle-time, bottleneck, rework, and scorecard metrics

MetricsService gains four read methods on the audit_log + tasks data: per-stage
cycle time reconstructed from the transition journey (excluding the named
qa_fail/pr_fail events), bottleneck distribution (cumulative dwell + live parked
counts), rework rate (overall/by-team/by-agent with rejector attribution + cost
via spawn-session task_id), and a fused per-agent/per-cell scorecard. Dataclass
models with to_dict(). Verified against a real Postgres journey.

* feat(observability): cycle-time/bottleneck/rework/scorecard read endpoints

Thin read-only routes on the dashboard router delegating to MetricsService:
/metrics/cycle-time, /metrics/bottlenecks, /metrics/rework, and
/metrics/scorecard/{agent,team}. 404 when an agent scorecard target is absent.
5 route tests (200 + shape + the agent-404 case).

* feat(panel): Delivery observability tab (cycle-time, bottlenecks, rework, scorecards)

A third Metrics tab built on the observability endpoints: a per-stage
cycle-time bar chart, a bottleneck panel (worst stage + cumulative dwell +
live parked counts), a rework panel (rate + by-team + by-agent attribution +
cost), and per-cell scorecards. Reuses Recharts + Card/Badge/Skeleton and the
React-Query hook pattern; observabilityApi mirrors usageApi with mock-mode
fallbacks. tsc + eslint clean; 113 panel tests pass.

* docs(observability): changelog + CLAUDE.md for the delivery dashboards

* feat(gateway-health): recover a broken-but-alive agent instead of protecting it

The verb-heartbeat cannot tell a quiet-healthy agent from one whose MCP gateway
is broken (a corrupted /app/.venv firing no verb) yet whose container is up — the
reaper's live-skip would shield it forever. The reaper now probes the gateway
out-of-band (docker exec: does the gateway venv import its deps?) and, once it
has been broken past gateway_health_grace_seconds (tolerating a transient probe
miss), kills + evicts the container so it falls through to release + respawn.
Probe-inconclusive or healthy spares the container. Gated by
gateway_health_enabled (default-on reliability fix; in the panel Feature Flags).
Defers the optional agent-side self-check + full registry re-adoption — the
reaper's docker-liveness fallback already recovers a broken-after-restart agent.

* docs(gateway-health): changelog + CLAUDE.md for broken-but-alive recovery

* docs(observability): user-facing docs for the Delivery dashboards + gateway-health

Documents the new Metrics -> Delivery tab (cycle-time, bottlenecks, rework with
rejector attribution, cell scorecards) in the panel guide and the operations
health-and-metrics guide, and adds the gateway-health env vars + an agent-gateway
recovery note. Published MkDocs site only; settings.md's default-off flag table
intentionally omits the default-on gateway-health flag (same as overload-break).

* chore(release): cut 0.10.0 (changelog section + version refs)

* fix(gateway): exempt PM coordinators from single-task claim guards

A Main/Cell PM plans and delegates many root tasks in parallel; the work
then runs in the delegated cells, not in the PM's own hands. But the
claim-time concurrency guards meant for developers — already_active and
paused (the latter firing after i_am_idle auto-pauses the PM's own
umbrella) — were applied to the PM too, so once it held one root it could
never plan a second: it thrashed between its claimed roots and respawned
forever, burning tokens for zero progress.

_run_claim_guards now skips already_active/paused for the coordinator PM
roles (_COORDINATOR_ROLES = {main_pm, cell_pm}); only unmet_dependency — a
real upstream sequence constraint, which parks the root back to pending —
still gates a PM. paused_tasks_guard also excludes the target task itself,
so a PM re-entering its own paused umbrella never self-blocks.

Tests: a coordinator plans a second root with one in_progress + one paused
sibling (full path + claimed-recovery path), the paused target exclusion,
and the developer guards still fire. Repurposed the pre-fix test that
asserted the now-removed PM block.

* fix(metrics): coerce SQL avg/extract hours aggregates to float (panel toFixed crash)

EXTRACT(epoch ...) returns numeric on PostgreSQL 14+, which asyncpg surfaces
as a Decimal; a Decimal serializes to a quoted JSON string, so the panel's
avg_cycle_hours.toFixed(1) (and the other hours fields) threw 'toFixed is not
a function' and blanked the Delivery tab.

A single _as_hours helper now rounds every SQL-averaged hours field to a real
float — avg_cycle_hours on the new scorecards plus the pre-existing
avg_completion_hours / avg_blocked_hours / longest_blocked_hours. Token and
cost fields were already float()-cast and are unaffected.

Regression test asserts _as_hours coerces Decimal -> float and preserves the
None/zero behavior.

* feat(panel): edit a task's sequence from the details page

A task's sequence (order within siblings, lower runs first) was display-only
with no way to change it from the UI, and TaskUpdate didn't carry the field
so PATCH couldn't set it either. The details page's Dependencies tab now has
an inline sequence editor mirroring the parent / dependency editors, and
PATCH /tasks/{id} accepts a sequence field (owner or privileged role) through
the existing generic update path.

* fix(mypy): green the full make-quality type gate

make quality runs 'mypy roboco/ tests/', which the per-module checks on the
0.10.0 branch never exercised. Two issues surfaced:

- The coordinator-exemption change added role_str to
  Choreographer._run_claim_guards but not to the ChoreographerHelpers
  protocol base, so the composed Choreographer had incompatible base-class
  signatures. Sync the protocol signature.

- The gateway-health / stale-reaper tests stubbed methods by direct
  assignment (orch._m = AsyncMock()) and typed their duck-typed task doubles
  as object, tripping method-assign / assignment / attr-defined. Switch to
  monkeypatch.setattr (keeping a local mock ref for the assertions) and type
  the doubles as Any — no type: ignore.

Full mypy roboco/ tests/ clean (785 files); the 21 runtime tests pass.

* fix(metrics): static cycle-time SQL — clear bandit B608 (CI gate)

The cycle-time query interpolated an optional team clause into the text() SQL
via an f-string, which bandit flags as B608 (hardcoded SQL) and turned the
merge gate red. The team value was always a bound parameter, so it was a false
positive — but the f-string is the trigger. Rebuilt as one static query with
(CAST(:team AS text) IS NULL OR a.details->>'team' = :team) and an always-bound
team param (CAST, not ::text — SQLAlchemy's :param parser collides with
PostgreSQL's :: cast operator, which broke the query as a stray param).

Full make quality green vs a real pgvector PG (all 21 gate steps).

---------

Co-authored-by: Renn F <rennf93@users.noreply.github.com>

CLAUDE.md

@@ -480,7 +480,10 @@ Server-side events reach these sockets through `roboco/api/websocket_bridge.py`,
 
 - **Provider rate limits** are tracked in Redis (`RateLimitStateTracker`, `roboco/services/gateway/`). On a provider 429 an agent calls `i_am_blocked(reason="rate_limited")`; the spawn gate then **queues** (never drops) further work for that provider, and a background probe-and-resume loop in the orchestrator clears the limit and revives parked agents when it lifts.
 - **Provider overloads** reuse the same park-and-probe break. A persistent model-API overload (HTTP 529 / 500 / 503 — the SDK already retries transient ones) parks the provider exactly like a 429 instead of crash-retrying the agent straight back into the overload and burning tokens; the overload is detected orchestrator-side from the dead container's log markers, and the background loop revives the parked work when it recovers. Gated by `ROBOCO_OVERLOAD_BREAK_ENABLED` (default-on).
+- **Gateway-health recovery** closes a blind spot in the stale-claim reaper: the heartbeat is bumped only by gateway verbs, so a broken-but-alive agent (a corrupted `/app/.venv` so no gateway tool imports) goes heartbeat-stale yet keeps its container up, and the reaper's live-skip would protect it forever. On a stale-heartbeat live container the reaper now probes the gateway out-of-band (`_probe_gateway_health` → `docker exec` the gateway venv imports) and, once broken past `ROBOCO_GATEWAY_HEALTH_GRACE_SECONDS` (a transient probe miss is tolerated), kills + evicts it (`_maybe_recover_broken_gateway`) so it falls through to release + respawn; healthy or inconclusive probes spare it. Gated by `ROBOCO_GATEWAY_HEALTH_ENABLED` (default-on). It is the third leg beside the shipped bash-guard `/app` block (prevents the self-corruption) and the reaper Docker-liveness fallback (stops over-reaping live containers).
+- **PM coordinator concurrency.** A Main / Cell PM plans and delegates many root tasks in parallel — the actual work then runs in the delegated children/cells, not in the PM's own hands. The claim-time concurrency guards that keep a *developer* to one task at a time (`already_active` / `paused`, in `roboco/services/gateway/claim_guards.py`) are therefore **skipped for the coordinator PM roles** (`_COORDINATOR_ROLES = {main_pm, cell_pm}`, consulted in `_run_claim_guards`); only a genuine upstream **sequence dependency** (`unmet_dependency`, which parks the task back to `pending`) holds a PM's root back. Without this a single PM that claimed one root could never plan a second — it thrashed between its claimed roots and respawned forever, burning tokens for zero progress (the live `i_am_idle`-auto-paused-umbrella deadlock). The `paused` guard also excludes the target task itself, so a PM re-entering its own paused umbrella never self-blocks.
 - **Token usage** is captured per agent session from the Claude Code transcript via the SDK server's `/usage/sync` (hook → orchestrator finalize → `agent_spawn_sessions` → `daily_usage_rollups` → dashboard). Cost uses provider-aware pricing in `roboco/billing/pricing.py` (Anthropic priced; local/Ollama intentionally `$0`). The token sweep also publishes `USAGE_SNAPSHOT` to `/ws/system`, so the dashboard's "Token Usage & Cost" panel updates live and falls back to HTTP polling when the stream is down.
+- **Delivery observability** (the panel's Metrics → "Delivery" tab) shows how work *flows*, computed by `MetricsService` from data already captured — no new feature flag. Per-stage cycle time and the bottleneck distribution are reconstructed from the `audit_log` transition journey (each generic `task.<status>` event marks entry into a status; the named `task.qa_fail`/`task.pr_fail` events are excluded from the reconstruction). Rework rate reads `tasks.revision_count` — incremented once per transition into `needs_revision` at the single chokepoint `TaskService._emit_status_transition_audit` — and attributes each bounce to the QA / PR-reviewer via those named audit events; rework cost joins `agent_spawn_sessions.task_id`. Read-only endpoints: `/dashboard/metrics/{cycle-time,bottlenecks,rework,scorecard/agent/{id},scorecard/team/{team}}`.
 
 ### Startup Sequence