feat(review): file triage stage — classify changed files to drive packing order and prompt selection [v0.8]

[devops-thiago]

Problem Statement

Batch packing is size-based only: DiffBudgetPlanner orders files by additions+deletions descending and packs First-Fit-Decreasing (DiffBudgetPlanner.java:609-736). Importance never enters the picture, so when the spend ceiling (REVIEW_MAX_TOKENS_PER_REVIEW), the call cap, or a batch failure forces degradation, what gets skipped is whatever happened to pack last — possibly the auth change, while a giant generated file got a full-budget batch.

Proposed Solution

Add a triage stage between diff load and budget planning that classifies every changed file:

• Deterministic first (free): category from path/content signals — source / test / config / IaC / docs / generated / lockfile — plus coarse risk signals (security-sensitive paths, public API surface, migration files). Most files classify without any model call.
• Optional cheap LLM call (concise binding, ledger-gated, counted in maxAiCalls accounting): rank only the ambiguous remainder by review priority. One small call with file paths + hunk headers, not diffs — predictable, tiny prompt. Config-gated, default off.

Consumers of the classification:

1. Packing order: importance-first (risk tier, then size) instead of size-only, so degradation always sheds the lowest-value files. The existing disclosure machinery (omittedFiles, spendCeilingSkippedFiles, withheld-material notice) keeps reporting honestly.
2. Dimension routing input: batch composition by category makes per-batch dimension selection (the v0.7 routing issue) sharper — e.g. pack config/IaC files together so only those batches carry IaC rules.
3. Lightweight lane: docs/generated/lockfile tiers can share a minimal sanity-check prompt or be summarized-only, rather than consuming full-prompt budget.
4. Model routing: the risk tier is exactly the signal feat(ai): risk-based model routing and fallback #43 (risk-based model routing) needs — this issue provides the classifier, feat(ai): risk-based model routing and fallback #43 consumes it.

Touchpoints

New review/FileTriage collaborator; review/DiffBudgetPlanner.java (ordering + lane assignment), review/FindingPipeline.java, config/ThrillhouseConfig.java (gate + tier globs, extensible via .github/thrillhousebot.yml like review.ignored-files).

Acceptance criteria

• Deterministic classifier covered by unit tests over representative path sets; LLM ranking optional and fail-open (falls back to deterministic order).
• Under a forced spend ceiling, high-risk files are demonstrably reviewed before low-risk ones (test via DiffBudgetPlanner plan inspection).
• No change in behavior when the feature gate is off.

Alternatives Considered

• Pure LLM classification — unpredictable cost and an extra call on every review; deterministic-first keeps the common case free.