feat(impeccable-skills-reviewer): reduce token bloat, fix retry rate, add skill-selection sub-agent

[Copilot]

The Impeccable Skills Reviewer had a ~65% retry rate, 44K–50K output tokens on large PRs, and a max-continuations: 6 ceiling that combined to push runs over the 15-minute timeout.

Changes

Retry rate & timeout

• timeout-minutes: 15 → 20 to reduce timeout-induced retries
• max-continuations: 6 → 3 — sufficient for the read → classify → review → submit flow
• Added Log retry attempt pre-agent-step: emits a diagnostic when run_attempt > 1

Output token constraints (## Review Constraints)

- Keep each inline comment under 120 words. Put extended reasoning in a `<details>` block.
- If the PR diff exceeds 1000 lines, limit inline comments to 5 maximum.

Skills discovery pre-fetch

New Build skills manifest pre-agent-step concatenates all installed SKILL.md files into /tmp/gh-aw/agent/skills-manifest.txt (paths deduplicated via sort -u) before the agent starts — eliminates the runtime find tool call.

select-skills sub-agent

New inline sub-agent (model: claude-haiku-4.5) replaces the inline skill-discovery step in ## Process. It reads the pre-fetched PR data and skills manifest and returns a JSON array of the 1–3 most relevant SKILL.md paths. Classification runs on a cheap model; the primary sonnet agent goes straight to reviewing.