feat(eval): golden-task eval harness + doctrine cohort stamp (#655)

* fix(notifications): exponential backoff + CAS claim for expired-unacked re-escalation

The sweep re-escalated every expired unacked ack-required notification
on every ~60s tick, forever — the live incident: 3 fresh blocker
escalations + Telegram DMs per minute from a static stale pile. Now
each notification carries reescalation_count / last_reescalated_at /
reescalation_delivered_count (migration 079): first fire at expiry,
then doubling intervals from 1h capped at 24h, hard stop after
ROBOCO_NOTIFICATION_MAX_REESCALATIONS (default 5) with one permanent
log carrying attempts-vs-delivered so 'seen and ignored' is
distinguishable from 'route never worked'. The due/wait/capped decision
is a pure function in foundation/policy/communications.py.

Per adversarial review, the attempt slot is claimed by compare-and-set
(UPDATE ... WHERE reescalation_count = :n) BEFORE delivery — the
previous draft leaned on the 60s dedup window, which never engages for
BLOCKER_ESCALATION (_LOOP_PRONE_TYPES excludes it), so concurrent
sweeps would have double-delivered. A lost claim skips delivery
outright. Legacy rows read as count=0 and keep today's first-fire
semantics. 61 tests incl. a two-session CAS race and a real alembic
upgrade/downgrade round trip.

* feat(budgets): per-task and per-project cost budgets (flag-gated)

tasks.budget_usd + projects.monthly_budget_usd (migration 080, chained
on 079; adds ix_agent_spawn_sessions_task_id since both enforcement
seams filter on bare task_id). Behind ROBOCO_TASK_BUDGETS_ENABLED
(default off, feature-flags card) — verifiably inert when off.

Claim-time: a project-month-spend guard applies to WORK-STARTING claims
only (i_will_work_on / i_will_plan) — per adversarial review, review/
doc/gate/inbound-PR claims are exempt so in-flight work can always
finish reviewing and merging at cap. Spend counts closed sessions'
estimated_cost_usd PLUS open sessions priced live from token snapshots
(the original closed-only sum read parallel long sessions as $0).

Sweep-side: the existing budget sweep also prices the active task's
spend vs budget_usd (TaskType defaults when null); on breach the task
is BLOCKED (HUMAN resolver, budget marker) BEFORE the graceful stop so
the unclaim no-ops and the dispatcher never respawns onto it, and the
CEO notification names both recovery steps. unblock on a budget-blocked
task re-checks live spend and refuses while still over — no silent
re-breach loop. Panel: budget inputs in both dialogs (0 rejected — a
zero budget silently blocks everything), spend logic consolidated in
TaskService.task_spend_usd. 42 new tests incl. a real-DB spend-query
suite and a two-tick non-refire sweep test.

* feat(eval): golden-task eval harness + doctrine cohort stamp

roboco/eval: 6 BenchTaskSpec fixtures run through the real lifecycle in
a disposable environment (the e2e_smoke harness's fake GitHub + local
git origin + throwaway DB catalog — real isolation, not convention),
scored deterministically (terminal status, revision_count, cycle time,
tokens/cost via the agent_spawn_sessions task_id join) plus a local-
model judge whose output is nested under a non_deterministic-marked
object so cohort diffs don't read judge noise as regression. CLI:
python -m roboco.eval run --role <slug> --cohort <name>. Source-
checkout-only by declared posture (deptry-scoped ignore + a hard
ImportError guard naming why; tests/ never ships in images or wheels).

agent_spawn_sessions.doctrine_version (migration 081, chained on 080)
is stamped at spawn-session finalize from the composed prompt layers —
with the session's model column it identifies a cohort durably.

Per adversarial review: bench runs patch the vault flags off (they were
writing real markdown into the operator's vault), and the real-spawn
OrchestratorStageSpawner is deliberately cut to NotImplementedError —
spawned containers' MCP wiring resolves to the production orchestrator
under real agent UUIDs, so real spawns wait for a dedicated follow-up;
the injectable scripted spawner is the working path. Full suite 13852
passed / 94% coverage in the source worktree; deptry/mypy/xenon clean.

---------

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

pyproject.toml

@@ -198,6 +198,13 @@ select = [
 # cost; ARG001 covers FastAPI path params the fake-GitHub handlers must
 # name but not read.
 "tests/e2e_smoke/*.py" = ["PLC0415", "ARG001"]
+# The eval bench (offline CLI, not part of the served app) defers heavy/
+# optional imports — tests.e2e_smoke.harness, roboco.runtime.orchestrator,
+# roboco.services.task, asyncpg — to call time for the same reason
+# tests/e2e_smoke and roboco/services do; PLR0913 covers the stage-driving
+# and scoring call surfaces (task/spawner/role/timeout tuples), same
+# rationale as roboco/services/gateway/**.
+"roboco/eval/*.py" = ["PLC0415", "PLR0913"]
 # PTH119: _grok_usage_json sanitizes the agent id with os.path.basename — the
 # path-injection sanitizer CodeQL's query models; the pathlib equivalent
 # (Path(...).name) is not recognized by that query, so we keep os.path here.
@@ -233,10 +240,18 @@ select = [
 # signature for keyword-argument compatibility (mypy override check), but the
 # stub bodies are empty — ARG002 would require renaming them, which breaks mypy.
 "tests/unit/services/test_optimal_grounding.py" = ["ARG002"]
+# _FakeJudge.score overrides BenchJudge.score — same override-signature
+# rationale as test_optimal_grounding.py above (a fixed fixture/diff/notes
+# stand-in body has nothing to do with those args).
+"tests/e2e_smoke/test_eval_bench.py" = ["ARG002"]
 # Collision-builder test helpers mirror the builder's many keyword inputs
 # (parent/project/intends/migration/shared/sequence) — bundling them would
 # hurt readability more than the arg count hurts.
 "tests/unit/gateway/test_collision_context.py" = ["PLR0913"]
+# _metrics()'s many optional kwargs mirror DeterministicMetrics' own field
+# count (a plain, no-defaults dataclass) — same rationale as the collision
+# builder above.
+"tests/unit/eval/test_scoring.py" = ["PLR0913"]
 
 # =============================================================================
 # MyPy Configuration
@@ -452,8 +467,15 @@ DEP002 = [
     "types-passlib",
     "types-PyYAML",
 ]
-# DEP003: Starlette is a transitive dep of FastAPI, but BaseHTTPMiddleware is needed
-DEP003 = ["starlette"]
+# DEP003: Starlette is a transitive dep of FastAPI, but BaseHTTPMiddleware is needed.
+# "tests": roboco/eval/runner.py deliberately imports tests.e2e_smoke.harness/arcs
+# (the offline eval bench's disposable-project machinery — see that module's
+# docstring) — deptry sees the local `tests` package as an unresolvable
+# transitive import since it isn't a PyPI dependency at all. The runtime side
+# of this is guarded separately (an ImportError there raises a clear "this
+# needs a source checkout" error), so this is a lint-posture ignore, not a
+# correctness gap.
+DEP003 = ["starlette", "tests"]
 
 [dependency-groups]
 dev = [