[Epic]: Review quality — deterministic severity and complete retrieval

### Summary

Successor to the closed epic **#252**. The product contract is the P1/P2 line and the claim that the reviewer read what mattered. Today severity flips on identical evidence, TSX repos index at ~2.5% symbol coverage, and the dashboard meant to size the context budget renders healthy and broken identically.

### Intent

Everything here decides whether a review can be **trusted**. A reviewer that returns a different severity for the same finding, or that never saw the component it is reviewing, is not a quality problem — it is a correctness problem wearing a quality label.

### Children

- **#285** — deterministic severity and stable anchoring
- **#623** — TSX is indexed as TypeScript, so JSX components produce no tree-sitter chunks
- **#600** — prompt-budget dashboard divides by an unguarded denominator
- **#639** — spike: summary-first block abstraction

### Acceptance Criteria

1. A golden set of ≥10 findings is **committed to this repo** as fixtures (diff + expected severity), re-reviewed K times at a **model pinned in the harness config**; zero same-evidence P1↔P2 flips across K runs.
2. A `.tsx` fixture directory produces ≥1 tree-sitter chunk per top-level JSX component (0 today) and **zero** `chunk_type: window` fallbacks.
3. The prompt-budget panel renders a distinguishable state when its denominator is zero.

### Out of Scope

- RFC-0002 layered indexing (#244) — larger, warrants its own epic once scoped.

### Technical Context

⚠️ **#285's current AC cannot be measured**: its golden set does not exist, its fixture is another team's PR that **merged 2026-07-10** (so the re-review cannot be re-run), and its model is a catalogue entry in `ai-helm-values` that gets swapped — ADR-0075 already consolidated GLM-5/5.1 → 5.2, which would re-roll severity with no code change here. AC 1 above replaces all three moving parts.

⚠️ **#623's defect is real and unfixed**: `chunker.rs` still maps `"typescript"` → `LANGUAGE_TYPESCRIPT` with no `LANGUAGE_TSX` anywhere. Only its *measurement* (an external repo, an unrecorded commit, "materially improved") was unstable.

### AI Usage Declaration

Drafted by Claude Opus 5 during the 2026-09-01 triage sweep. AI output is not truth; the owner verifies.