Hunt evidence review and continuity

Status: Implemented review/integrity tooling on the 2.3.0 integration branch; offline scorer and CI gate only, no live efficacy measurement.

This change builds on the recorded live acceptance in hunt-authorization-behavior.md at bbaa84c9, documented by 80abc5d0. Those two positive leads and the no-lead control remain acceptance evidence, not a human-productivity study or verified recall gain.

Retained history versus current assessment

Investigation responses now include candidate_history, candidate_relation, and candidate_matches_latest_attempt. The compatibility candidate field is the most recent retained association, even when a later attempt is denied or inconclusive. historical_only does not claim that the latest attempt reproduced the lead. The latest assessment, proof state, and immutable candidate assertions are unchanged. Propose/read/skip use the same non-executing history projection; none materializes a lead.

Candidate materialization locks the existing proposal row before checking its deterministic link. The candidate, observation, and link are committed together. Concurrent retry callers must recheck the link after acquiring that lock. An error writing the link rolls the candidate and observation back. No new table, proof path, or execution capability is added.

Review in the existing Hunt page

Opening a saved run at /hunt?run=<hunt UUID> displays an investigation review panel above the operational ledger. It pages existing proposals, reads one selected investigation, and shows current assessment, retained leads, unresolved questions, attempt/receipt/transaction references, and supported-request limitations. Refresh is explicit. The panel cannot propose, approve, execute, or promote anything; these changes do not broaden replay support.

The existing graph_nodes query accepts node_type and hunt_id filters. Its attributes projection exposes allowlisted UUIDs, digests, and enums for authorization records, not arbitrary graph attributes or raw request content. Target scoping and cursor/filter binding remain enforced. Unsupported attributes are marked omitted. This is bounded resume metadata, not a complete target graph.

Offline human-review measurements

The existing scorer accepts optional --investigations (an array of unmodified investigation GET responses) and --review (operator-recorded labels and time). Candidate associations must link to an executed action in the canonical Hunt record. Example review file:

{
  "hunt_id": "<the exported Hunt UUID>",
  "human_minutes": 12.5,
  "useful_lead_ids": ["<an exported candidate UUID>"],
  "not_useful_lead_ids": [],
  "duplicate_experiments": 0
}

These values are operator-recorded, not technical proof. Missing measurements are null, not zero. Labels must be disjoint and refer to exported candidates. They never increase verified findings or recall, and a single review cannot demonstrate productivity improvement. Exports remain operator-trusted; the scorer does not authenticate edited files or execute verification.

Settled conservative full-reservation charges are now reported as upper bounds, separately from measured traffic. Unknown accounting cannot establish complete exact cost.

Validation boundary

Hunt record integrity runs the focused Python tests, a disposable real PostgreSQL locking and rollback test, UI unit tests, and a UI build. Set HUNT_TEST_POSTGRES_DSN to a disposable local database to run the PostgreSQL test locally; configured-but-unavailable databases fail rather than silently skipping. Each test removes only its own random schema.

The database test executes production repository SQL but uses a minimal schema, not the full migration path. The checks do not run an offensive workload or establish live worker, browser workflow, multi-model, or paired human-productivity acceptance. The unchanged independent protocol in hunt-investigation-evaluation.md remains the efficacy evaluation requirement.