[kvlonge]

Why

Fresh-process warm compiles still repeated whole-project work despite valid fact caches. On the representative 3k semantic project, the unchanged path took 21.0s; SQL-test artifact handling alone rebuilt the full model dependency index once per test.

Changes

• Build SQL-test model closures once per compile and share them across artifact identities.
• Add v10 dependency-aware compact analysis generations with content/version identities, Merkle-style upstream keys, partial reuse after edits, per-entry fallback, integrity validation, two-generation retention, and corruption/contention-as-miss behavior.
• Report analysis batch hits, entry hits, misses, and bypasses in compile JSON.
• Document the target cache architecture and delivered vertical slice.
• Add 3k/5k/10k fresh-process CI guards covering cold, unchanged, leaf edit, macro edit, broad config invalidation, post-edit reuse, semantic fingerprints, cache size, and peak RSS under a 4 GiB cgroup.

Verification

• Focused compiler, target-writer, JSON, and seed-target tests: 131 passed.
• Sanctioned unit/integration run: 6,768 passed; 12 optional Dagster collection errors were resolved by installing all extras, then all 86 affected tests passed.
• uv sync --all-extras
• make check-ci
• Public-repository hygiene scan and independent bounded review: clean.
• Calibrated 3k: 5.99s unchanged, 6.17s leaf edit (about 400 MiB warm RSS).
• Calibrated 10k: 20.18s unchanged, 21.23s leaf edit (about 1.0 GiB warm RSS); cold/config peaks remained below 2 GiB.