[SizzleUnrlsd]

Summary

Phase 10 opens with the persistent cache of docs/architecture.md §11: a directory check keeps per-module results on disk and, on the next run, re-analyses only the modules whose inputs changed.

• Acceptance tests first (test: define persistent cache behavior), implementation follows.
• cache.py (root module, like engine): CachedModule (functions, summaries, call sites, findings), a JSON codec for summaries, external calls, call sites and findings, and ProjectCache, a directory of entries written atomically and read leniently: a corrupt, foreign or older-format entry is a miss. Entries are data, never code.
• Keys: a configuration fingerprint (engine, HIR, finding and plugin API versions; each loaded plugin's manifest and the digest of its code; the model table; the advisories; the dependency graph) combined with the module's identity and text, then with the keys of the project modules it imports transitively. A change anywhere below a module misses for that module; a change in a leaf leaves its importers cached.
• engine.analyze_project(..., cache=): cached modules never reach lowering. Their summaries seed the project index before the fixpoint, so incremental summaries come with the cache; their call sites are rebuilt into a CallGraph handed to ProjectContext, so the project plugins and the correlation keep working over a mix of fresh and cached modules. ProjectAnalysis exposes keys and reused.
• --cache DIR on the CLI, valid with --check on a directory.

Not in this PR, tracked on #41: incremental parsing (every file is still parsed each run, which is what the module graph needs), parallel analysis, memory eviction.

Refs #41