[kvlonge]

Why

A warm sqb compile with nothing changed still rediscovered, re-rendered and re-checked every model: about 8 s on a dense 3k project and 27 s at 10k. Checking that no input changed takes milliseconds (CHI-498, stage 1 of CHI-482).

Changes

• When no input changed, sqb compile replays the previous stdout (identical apart from compile_timings) and exit code, including failed compiles, and prints one stderr line.
• Inputs: project files (stat stamps with content-hash fallback), flags, selectors, --vars, target, cwd, SQLBUILD_*/SQB_* and template-read env vars, provider settings env vars, env files and secrets, Python, SQLBuild and native build identity, and loaded modules. target/compiled must match what this compile wrote.
• Full compile for --no-cache, --manifest, --debug, profiling, compile_cache = false, run.id, and SQLBUILD_DISABLE_COMPILE_REUSE=1. Docs list inputs and limits.

Verification

• No-change compile (fresh process): dense 3k 8.3 → 0.58 s, dense 10k 27 → ~1.5 s; cold and edit within noise.
• Real-CLI oracle per input kind (equal to --no-cache, miss exactly when it should), provider-settings, concurrent-writer, corruption and store-failure tests; compile and build E2E run once with reuse enabled.
• One independent review plus follow-up; make test, make check-ci.