[kvlonge]

Why

After a one-model edit, compile still rediscovered every declaration file, re-read every model's analysis from the cache and re-compared every compiled artifact, even when render reuse (#955) had already shown only one model changed (CHI-508).

Changes

• Discovery: non-model declaration files (sources, tests, macros, schemas) are reused when only existing model files changed; any other change rediscovers everything.
• Analysis: an unchanged model reuses its stored column analysis only when its render was reused, no other model shares its cache key, its cache key equals the one computed in this compile, and every direct parent's final signature in this compile matches the one recorded with it. Otherwise it takes the normal path. New analysis_reuse_hits counts analyses actually served.
• Writes: compiled files whose stored stamp and digest match are neither read nor rewritten.
• Semantic validation still runs in full.

Verification

• Real-CLI edit-chain oracle against --no-cache, including two-hop SELECT * chains with plan, build, another target or a no-reuse compile between edits, same-SQL twin models, and random chains; each new reuse fails the oracle when deliberately broken.
• Second edit after cold, alternating medians: dense 3k CPU 8.8 → 8.0 s (analysis ~1150 → ~435 ms), dense 10k and a ~1,400-model project also faster; cold and no-op unchanged.
• One independent review plus follow-up; make test, make check-ci.