[kvlonge]

Why

Editing one model recompiled every model from scratch. Compile reuse (#951) only helps when nothing changed. This change reuses unaffected renders after model-only edits, with output byte-identical to --no-cache.

Changes

Invalidation by input kind (change set C comes from stored file stamps and digests; invocation, runtime, env, settings and modules must be unchanged):

• Edits to existing model files only (C within the stored model paths, model path list identical): the edited models are rendered again; all others reuse the stored render, and its diagnostics and env reads are replayed. Renders that read run.id or provider settings are never stored.
• Shape propagation, semantics, contracts, built-in and custom rules, and artifacts run on full inputs through their existing per-key caches.
• Sources, tests, scenarios, singular audits, and source/seed audits are reused as whole groups. A model's attached audits are reused only with that model's render.
• Added, removed or renamed models, and macros, imported modules, config, path_defaults, declarations, scopes, seeds, sources, tests or audits: full render.

Storage: renders are stored only when a compile entry already exists, so cold compiles pay nothing. Files are a zlib-compressed base plus a bounded overlay, two newest kept, verified on read; any fault means a full render. Stored bytes are released once renders are decoded.

Verification

• Real-CLI DuckDB oracle: authored edit sequences plus seeded random chains, each step byte-compared with --no-cache; fails when change detection is broken. Save-policy test: cold stores no renders, the next miss does.
• One independent review plus follow-up; make test, make check-ci.
• Medians of 3 alternating runs at load < 3, dense 3k / 10k: cold 16.2 / 49.6 s, unchanged. First edit after cold +0.7 / +2.3 s. Later edits 8.6 → 7.6 s and 27.0 → 23.7 s, with peak RSS +0.1 / +0.35 GB. Render store 5 / 18 MB.