feat(sdk): surface verifierRejections across reports, CLI output, and findings-file export

[CalebKAston]

Summary

A finding the verification pass rejects simply disappears today — no count, no reason, visible nowhere in Warden's output. This adds a verifierRejections: {count, reasons[]} field, computed in verifyFindings() and threaded through postProcessFindings() onto SkillReport.

It's surfaced everywhere failedHunks/failedExtractions already are: the terminal report, the Reporter's aggregate summary (log + TTY modes), the JSONL run log (per-chunk and run-summary), and the GitHub Action's findings-file export. The run-level aggregate (JSONL summary, Reporter, terminal) intentionally surfaces only a count, not the reasons — matching the existing failedHunks/failedExtractions precedent. Reasons stay available at the per-skill chunk-record level.

Real-world verification

This has been running in production for a period feeding an internal data pipeline into an observability platform for tracking skill health. Querying that platform:

• ~3,900 skill runs recorded in the window.
• 555 of those runs (about 14%) logged at least one verifier rejection, totaling 874 rejected findings across three different review skills.
• Spot-checking a sample of the recorded reasons shows the verifier doing real, substantive work, not rubber-stamping: the large majority cite a finding anchored outside the PR's actual changed files (out-of-scope), or reason through why a cited convention doesn't actually apply to the specific pattern flagged — exactly the "would have been noise for the author" cases this field exists to make visible.

Test plan

• pnpm lint && pnpm build && pnpm test — all green (93 test files, 1839 passing, 4 pre-existing skips)
• verify.test.ts — rejection counting, default-reason fallback, multi-finding aggregation
• output.test.ts — findings-file export present/absent
• terminal.test.ts, reporter.test.ts — per-skill and aggregate warning lines
• jsonl.test.ts, main.test.ts — schema round-trip through the JSONL chunk stream and run summary
• tasks.test.ts — end-to-end through the CLI task runner