Add persistent cache and incremental parallel analysis #41

Cost model: Ttotal = Tcore + Tunique_analyses + Σ Tplugin, never P × Tcore.

• Persistent on-disk cache keyed by IR fingerprint and analysis/plugin versions (feat: persistent cache #42: cache.py, keys from source, engine/schema/plugin versions, plugin code, models, advisories, dependency graph and transitive import keys; --cache DIR)
• Incremental parsing with Tree-sitter (retired by decision on 2026-09-04, docs: frontend decision #50: the frontend stays the stdlib ast adapter behind build_hir; the per-module persistent cache covers incremental analysis at this scale; decision note in docs/architecture.md §3.1)
• Incremental summaries (feat: persistent cache #42: cached summaries seed the project index, only changed modules and their importers are re-analysed)
• Parallel analysis over the DAG at module granularity (feat: parallel analysis and memory eviction #43: ModuleGraph.schedule waves of strongly connected components, --jobs N process pool; function-level parallelism inside a module is not needed while summaries are per module)
• Memory eviction of PyIR and results once summaries exist (feat: parallel analysis and memory eviction #43: ResultsEvicted pass keeps semantic tables and engine inputs only)

Reference: docs/architecture.md §11 Cache, §29 Performance Model, §30 Memory Complexity.

Delivery: #42 (cache and incremental summaries)
Delivery: #43 (scheduling, parallel analysis, eviction)
Delivery: #50 (decision note, Tree-sitter retired)