Paired with the enterprise PR that rewires the offline playbook tuner onto a single per-playbook impact signal. Merge this first, then the enterprise repo pins it.

Why

The tuner now reads the per-playbook impact verdict in retrieved_learning_evaluation and nothing else. Two consequences land in OSS:

1. The retrieval-capture subsystem becomes unreferenced — it existed to record what was served, which Interaction.retrieved_learnings now carries.
2. The tuner needs two things OSS didn't offer: coverage it can buy without side effects, and a windowed read that doesn't silently truncate.

What's here

Independent per-family sampling

New AgentSuccessConfig.retrieved_learning_sampling_rate, default None = inherit sampling_rate — an org that never opts in keeps exactly its previous behavior and cost.

New agent_success_evaluation/sampling.py: the publish gate samples both families once, schedules when either admits the session, and passes two booleans to the runner. One computation site, so the families cannot diverge. A session sampled only for retrieved-learning never pays the session-success judge — which is the point: it lets an operator buy coverage without also raising the agent-success bill.

Direct callers (regen jobs, on-demand grade) run both families as before. Sampling is a scheduling decision, not a runner one.

degraded is settled, not broken

failed/pending committed nothing and had no retry trigger, so such a session was invisible forever. The publish callback now re-arms it on the existing scheduler, bounded to 3 attempts.

degraded is deliberately NOT retried. It is an applied, fingerprint-fenced commit: the rows are persisted and only the failed chunks carry NULL impact. Retrying it re-executes every relevance+impact chunk and delete/re-inserts already-committed rows — and a deterministically degrading chunk (over-length learning, content-filter refusal) degrades again on every attempt, so a bounded sweep burned 4× the judge bill and bought nothing.

Instead: new SETTLED_RETRIEVED_STATUSES (= TERMINAL | {degraded}), exposed via an optional statuses arg on the terminal-state fence, so consumers read those rows directly. Contract-tested that widening the set does not weaken the freshness check — a degraded session whose transcript changed is still rejected.

⚠️ Best-effort retry: the attempt counter and schedule are in-process (GroupEvaluationScheduler is an in-memory heap), so a deploy or task recycle inside the delay window drops pending retries, and with N tasks a session can start N independent chains. Durability needs a persisted queue — not done here.

Exhaustive windowed read

get_retrieved_learning_evaluation_results_in_window() on the base + SQLite. The existing paged reader hard-caps at limit=100 with no cursor; a consumer that treats the window as the complete set must not inherit that silent truncation. Contract-tested at 0 / 100 / 101 / 250 rows.

Retrieval-log retirement

Deletes PlaybookRetrievalLog / PlaybookRetrievalLogItem, RetrievalLogMixin, and the configure_retrieval_capture_hook search hook.

Deliberately KEPT:

• the AuditEntityType "playbook_retrieval_log" literal — historical audit_events rows still carry it, and removing it makes Pydantic reject them on read;
• the retention targets + cascade — the tables still hold PII until a later release drops them. Removing a purge before dropping its table strands PII.

Verification

• 4613 passed; the 10 failures are pre-existing sqlite-vec/FTS ordering flakes, confirmed by running the same suite on clean main.
• Ruff + Pyright clean.
• The sampling-default test now constructs the config without the field. The previous version passed None explicitly, so it asserted its own argument — which is exactly how a 0.1 default drifted in unnoticed. Verified the new test goes red if the drift is reintroduced.

Summary by CodeRabbit

• New Features
  • Added independent, deterministic sampling for agent-success vs retrieved-learning evaluations, with a new optional retrieved_learning_sampling_rate config (inherits existing sampling when unset).
  • Added inclusive time-window reads for retrieved-learning evaluation results and bounded retry scheduling for pending/failed retrieved-learning outcomes.
• Bug Fixes
  • Preserved read compatibility for existing audit records containing legacy request audit entity types.
• Documentation
  • Clarified retained behavior for playbook_retrieval_logs during the retirement period.
• Refactor
  • Removed retrieval-log models and retrieval-capture hook APIs.
• Tests
  • Expanded per-family sampling, window-read, and retry coverage; removed obsolete retrieval-log and unified-search capture-hook tests.