Findings: leads, one eval's reviewed findings as hypotheses for the agents

[MattFisher]

Roadmap milestone 1, third slice, per the plan. Targets feat/findings.

What it is. inspect-audit-findings leads --out DIR [--sample ID] [--write PATH] EVAL renders one eval's current findings as LEADS.md: hypotheses with a location attached, for an agent to confirm or retire. Same current view as the summaries, review files applied. Suppressed observations dropped and counted; accepted issues listed first; groups capped at three examples with a pointer to the run file for the rest; skipped checks listed with one-line reasons. Every example carries its record id, which is what the agent cites when a register finding rests on a lead. --sample narrows to observations naming that sample (dataset and transcript locations) and says how many eval-wide observations were left out.

Staging contract. docs/leads.md describes the hooks: the investigator stages /inputs/findings/LEADS.md plus the run files and names it in seed.json; the sample auditor stages /audit/leads.md beside discrepancies.md; the chess import later reads cited record ids back into status history. Those hooks touch James's code and prompts and go as a separate PR. This module imports nothing from them.

Also. Two renderer helpers (inputs_lines, sorted_findings) are now public and shared. Under lint 0.9's rules mode, skipped rules keep their reason from the matching outcomes row; the trial had shown "Not examined" listing bare codes. The lint adapter tolerates outcomes: null.

Trial. StereoSet with lint and dataset producers and one accepted issue: preamble, Inputs, the issue, two duplicate-question groups capped at three with record ids, ten skipped lint rules with reasons. --sample <id> narrowed to the one observation on that sample.

Review. A fresh reviewer found one Important issue (--sample reported the eval's suppressed count as the sample's and hid eval-wide leads silently) and nine minors; the Important one and five minors that change what the agent reads are fixed with tests that failed first. The rest are in the first comment.

Gate: pre-commit clean, basedpyright 0, 535 passed.

🤖 Generated with Claude Code

[MattFisher]

Deferred minors from the branch review, none blocking:

• Within a group, examples are ordered by location, so observations already linked to an issue can take all three example slots while new ones fall into "and N more". Sort unlinked observations first, or say "k of n linked to ISS-x" on the heading.
• The run-file pointers (runs/<id>.run.json) are relative to <out>/<slug>/. The investigator staging copies that directory beside LEADS.md; the sample-auditor staging in docs/leads.md copies only leads.md. Either copy the runs too or mark the pointers as reference only in that contract.
• --write pointed at a directory is a traceback rather than exit 2.
• The no-runs error could list the evals present in the current view, since stereoset without the inspect_evals/ prefix is an easy mistake.

Left as they are, on purpose: --sample does not match LogLocation.location_hint (no producer puts a sample reference there); retracted findings are not filtered out yet (nothing produces that status; the loop-closing slice adds it); the per-sample auditor gets only leads that name its sample, now with a count of what was left out.

Posted by Claude Code on Matt's behalf.