repo-fleet-hygiene: 869 findings across 6204 lines with no rollup, and one target emitting multiple independent findings

[kyle-sexton]

Context: #2597

Problem

The report's presentation contract — repository sections followed by five grouped confidence lists — does not survive contact with a real fleet. It produces a per-item enumeration of what is, operationally, one decision per repository, and it emits the same target more than once when several findings apply to it.

Evidence

One run: 869 findings, 6204 lines, 19 repositories. The bulk categories are 263 reclaimable-worktree, 214 merged-local-branch, and 163 merged-worktree — 74% of all findings in three categories that a reader will act on wholesale, per repository, or not at all.

Duplicate targets are real. …/github-iac/.claude/worktrees/feat+176-work-class-label-axis emits both HIGH locked-worktree and MEDIUM worktree-nested-in-repository. A reader triages the same path twice and, because the two land in different confidence groups, may not notice they are the same path.

The grouped-list contract also has no way to say "this repository is clean" — the question an operator actually asks of a hygiene tool. It answers "here are 869 observations."

Proposed change

• Aggregate first. Default output is a per-repository rollup table with counts by finding kind, plus a fleet header. Per-item detail moves behind --detail or into a sidecar file.
• Group findings by target. One entry per path or branch, carrying every applicable finding, rather than one entry per (finding, target) pair.
• Emit a per-repository verdict line — CLEAN / N candidates / BLOCKED (evidence gap) — and a fleet verdict at the top.
• Emit machine-readable JSON alongside the human report, so downstream skills and the execution half consume structure instead of parsing prose. This is also the manifest referenced in repo-fleet-hygiene: the plugin reports per-repository facts a sibling already owns, and cannot perform the fleet-wide cleanup it exists for #2597.

The existing rule that a truncation or evidence gap must be reported per repository rather than once globally should be preserved in the rollup: a reader skimming a summary table must still see that a given repository's absent findings prove nothing.

Acceptance criteria

• A 19-repository fleet produces a report that fits on a screen by default, with detail available on request.
• No target appears in more than one finding entry.
• Every repository carries an explicit verdict, including "clean".
• A JSON artifact is produced and is sufficient input for the execution half.