coverage-report

# coverage-report Numbers come only from the tool. The model never estimates, infers, or rounds coverage. The model's only role is joining tool output to the manifest's reasons and narrating the observations section. Read `${CLAUDE_PLUGIN_ROOT}/rules/coverage-report.base.md` and the repo's `.claude/coverage/refs/coverage-manifest.yml` first. ## One command, any time The whole report is one wrapper — no arguments needed (it auto-detects the `.sln`): ``` ./.claude/coverage/tools/report.sh ``` It collects coverage, joins it against the manifest, and **writes a dated per-run folder `.claude/coverage/reports/<YYYY-MM-DD>/` containing `REPORT.md`, `REPORT.html`, and `CANNOT-TEST.md`**, then prints the Markdown. It exits non-zero only for a check named in the manifest's `gate.enforce` (empty by default, so a breach prints as `ADVISORY` and the script still succeeds); a failing build or test run is fatal regardless, from `run-coverage.sh`. Each run writes its own dated folder rather than overwriting a fixed path, so prior dates are preserved for comparison; override the date with `REPORT_DATE=YYYY-MM-DD` to re-emit under a specific day. `CANNOT-TEST.md` is generated from the manifest `cannot_test` entries (grouped by blocking construct, each row citing target, reason, and unlock), not hand-written. This is the *identical* join CI runs — the report a developer sees locally is the report on the PR. When invoked as the `coverage-report` skill, just run it and surface the result; d...