loop-evidencelisted

# Generate the Evidence Report The loop has a reviewed success candidate: verification passes, the success gate was reached (the implementer declared ready, or the harness forced the gate after consecutive MET stop evaluations), and, where review is enabled, the independent reviewer approved. Produce the report that lets a human review **evidence instead of diffs**. Report what the diff actually shows, not what the goal hoped for. The skill argument contains `baseline=<ref>` — the first commit for this task —, `logs=<path>` — the only task/run log namespace you may inspect —, `task=<task-id>`, and an `html=` token (`on`, `off`, or `auto`), which controls whether you ALSO write the HTML view (see "Write the HTML view" below). The argument may additionally carry `rejected='<reason>'`: a report generated earlier in this same gate was deterministically rejected for exactly that reason and deleted. Regenerate the FULL report, correcting precisely that defect — the other content requirements are unchanged, and the certification inputs (contract, checklist, manifest, observations) are frozen, so the fix is always in how the report cites them, never in editing them. At the **fleet integration gate** the argument additionally carries `merged=<id,id,...>` (every worker task that completed into this result — merged ones AND NO_OP ones, whose determination "verification passes with no changes needed" is evidence too) and `archives=.loop/docs/run-archive`. Then the report MUST cover e