feat(review): escalate review depth when large PR yields zero inline findings

[devops-thiago]

Problem Statement

Dogfood on v0.3.0 meta-PRs (#252, #268–#286): ThrillhouseBot posted 0–1 inline findings while maintainer ultra-review found ~25+ issues. The bot was a reliable CI gatekeeper but weak as a code reviewer on large refactors.

#335 fixes summary narrative accuracy; #53/#256 fixes truncation/coverage via map-reduce. This issue is the escalation heuristic: detect "suspiciously clean" large PRs and deepen the pass.

Proposed Solution

When changedFiles (or reviewable file count) exceeds a threshold and inline findings after verifier are zero (or all low confidence):

1. Trigger a second review mode — e.g. file-group map-reduce (spike(review): define a better large-diff handling strategy #53), /improve-style whole-PR pass (#316), or maintainer-configured "deep review" label
2. Append summary note: "Large PR with no inline findings — consider manual /review or /improve"
3. Optional: lower inline confidence threshold for large PRs (pairs with #105)

Touch points: ReviewOrchestrator, FindingPipeline, VerdictBuilder, config under thrillhousebot.review.*.

Related issues

• #53 — large-diff strategy (v0.4.0)
• #335 — summary narrative accuracy
• #316 — /improve command
• #105 — confidence gating
• #113 — eval corpus

Acceptance criteria

• Configurable threshold (files and/or LOC) triggers escalation path
• Zero-finding large PR gets measurably deeper second pass or explicit maintainer nudge in summary
• Small PRs unchanged (no false escalation noise)
• Tests with fixture PR over threshold + empty findings
• Documented in README config

Priority

Important — v0.6.0 with #335 and map-reduce landing.

Code of Conduct

• I have searched for existing issues and this is not a duplicate.