Queue Selection and Claim Lifecycle

This is the complete normative Step 2 contract for vault-ingress. Execute normalization before claiming, and preserve every generation, freshness, replay, and recovery rule below.

Fresh queue work uses claim schema v5. Before changing any talk status, queue-state.py claim snapshots one immutable adherence_baseline for the entire selected batch, stamps required_return_schema_version: 5, and copies that exact snapshot into every member claim. The snapshot uses the current catalog fingerprint and pattern-scoring schema, includes only current processed/processed_partial talks, and excludes every active-batch filename before inspecting its prior score. Its as_of equals the claim's canonical claimed_at; do not edit or recompute it after the claim is written.

• Run "{python_path}" "{speaker_toolkit_root}/skills/vault-ingress/scripts/queue-state.py" {vault_root}/tracking-database.json normalize once. This one copy-on-write command migrates legacy skipped_no_video/skipped_no_transcript states from the talk's usable source capabilities. The shared root-aware assessor must actually read and validate a local transcript/deck/PDF before it counts; a non-empty or missing path is not a capability. Declared remote video/slide acquisition remains eligible for a claim-time download. A transcript_source provenance label alone is not capability. Only a record with no verified local source or remote acquisition path becomes skipped_no_sources. The same command uses the shared exact-generation selector to requeue valid processed/processed_partial results that cannot enter the active pattern cohort. For scoring v5, exact generation identity is necessary but insufficient: the shared freshness assessor also re-hashes every persisted source-located artifact against the database's vault and configured source roots. Missing, replaced, or otherwise drifted evidence is excluded and requeued. Transcript freshness also revalidates the hash-bound quality policy against current owner/provider duration; source- identity drift cannot leave a previously accepted partial transcript in the cohort. Its normalizations report carries ordered reason codes, deterministic freshness details where applicable, and the stored reprocess_reason for each changed row. Malformed current-generation identity or score lanes reject the complete command without writing; inflight, pending, already-queued, and skipped records are not generation-normalized. A repeated successful normalization with no new drift leaves the DB bytes unchanged. Completed claims remain immutable evidence on the requeued talk and move to history normally when a later claim is created.
• Claim each exact batch through queue-state.py ... claim --run-id <stable-run> --batch-id <stable-batch> --now <timezone-aware-ISO>. The command selects pending, needs-reprocessing, and retryable download failures, writes an atomic lease, increments reprocess_generation, and returns the claimed talk list. Every fresh claim is schema v5 and requires return schema v5. Never change a record to reprocessing-inflight by hand.
• Before claiming a non-YouTube talk whose transcript will be used as evidence, acquire the transcript and register its canonical vault-relative transcript_path through the guarded source-repair workflow. A worker-returned path cannot grant citation authority to that same return.
• Repeating the same live/completed run and batch is an idempotent replay: use the exact stored claims and baseline, and do not rewrite the DB. Recover an expired or stranded lease with queue-state.py ... recover; recovery keeps the saved snapshot in claim history. A later claim creates a new generation and takes a fresh pre-mutation snapshot rather than editing or reusing the old generation's baseline.
• Set slide_source per the PPTX → PDF → video extraction → none hierarchy in SKILL.md. Mark "skipped_no_sources" only if preflight confirms that transcript, slide, and video sources are all unavailable.
• If $ARGUMENTS specifies a talk filename or title, process ONLY that one.