feat(review,audit): deterministic coverage and cited-line verification

[iagogfe]

What

Two additions to ponytail-review and ponytail-audit, both mechanical, zero new scope:

1. Deterministic coverage. The file list comes from git diff --stat (review) or git ls-files (audit), and every file ends in one of two states: reviewed, or skipped with a reason. Reports close with covered: N/M. Lazy means reading less per file, never silently reading fewer files.
2. Cited-line verification. Every path:line in a finding is checked with sed -n '<line>p' before it ships. A citation pointing at } is a finding the author never locates.

Why

On larger diffs the model quietly cuts corners: it picks the "interesting" files and the report reads as if everything was covered. And line references drift. Both are shape problems, not judgment problems, so the fix is a contract on the output, not more prose asking the model to be careful.

Measured

Same model, same 12-file Rust diff, same prompt, one run each way:

	without	with
coverage	varied between runs, no declaration	12/12, per-file status
cited lines	1 of 6 pointed at a closing brace	3/4 exact, 1 in the right block
cost	86k tokens, 17 tool calls	75k tokens, 9 tool calls

The cost drop was unexpected but repeatable in my setup: enumerating first replaces exploratory wandering.

n=1 per arm, so take the numbers as direction, not proof. The shape guarantee (coverage table always present) is the part that held on every run I did.

Checks

• .openclaw/skills/ mirrors regenerated with scripts/build-openclaw-skills.js (the sync test caught my first push missing them).
• Local suite: 81/82 pass. The one failure (csv: correct pandas one-liner) reproduces on pristine main in my environment (no pandas installed), so it's environmental, not from this change.

Fit

Kept to the house style: recipe form, no prohibitions, a few lines per skill. net: stays the closing metric; covered: precedes it. No new dependencies.