refactor(categories): precompute section splits; benchmark follow-ups

[andreujuanc]

Summary

Follow-ups from reviewing the merged category Shapes/Review-criteria split (#40):

• Precompute category section splits at init. Each category .md is embedded and static, but the Shapes/Review getters were re-running splitSections on every call — ~4× per review on the hot path (once per GetShape/GetReviewCriteria per AOI, and again for the cache key vs. prompt build). Now each file is parsed once into parsedCategories at init(); the getters are map lookups. Prompt output and cache keys are byte-identical.
• Document the unmigrated GetCategoryShapes fallback as an intentional whole-file pass-through (surface everything rather than silently drop patterns).
• Make the deep-review benchmark's AOI scan provider-agnostic. It hardcoded Gemini and skipped without a Gemini key. It now follows cfg.FastModel (override via PRR_AOI_MODEL), pulls tuning from the model config, and skips only when a keyed provider lacks a key — so the benchmark runs keyless via claude-code.
• Record measured Opus 4.8 numbers on DefaultStrongModel (deep-review benchmark: 100% recall, 0 FP, 77% severity accuracy).
• Add init-drift and unmigrated-fallback tests.

Context

No behavior change to review/AOI prompts — the refactor preserves the recall-vs-precision split byte-for-byte (AOI scan gets Shapes only; deep review gets scoped Shapes + Review criteria). The benchmark change enabled validating the Opus 4.8 default keyless, which produced the figures now in the config.go comment.