[kvlonge]

Why

The compiler engine differential intermittently fails on CI because a warm compile misses the SQL-test artifact cache and re-plans its tests (CHI-570). The cache required an exact st_mtime_ns match, so any touch of an unchanged artifact, or a coarse timestamp tick on a slow runner, forced a miss even though the bytes were identical.

Changes

• Cache records now store a SHA-256 of the artifact contents.
• Matching size and mtime still reuse the artifact without reading it. When the size matches but the mtime moved, the artifact is reused only if its content digest is unchanged, and the record is refreshed with the new mtime.
• Cache version bumped from 3 to 4, so older cache files are ignored once and rebuilt.
• New unit test: an artifact whose mtime moved but whose bytes are unchanged is reused across two writes without re-planning. It fails without this change.
• No expectations or thresholds changed. The existing tampered-artifact test still re-plans when the bytes differ.

Verification

• test_compile_target_writer.py (15 passed), compile startup-imports test.
• ruff format and check, ty on the touched packages, fensu (0 faults), hygiene scan.

[kvlonge]

Held out of the merge queue: two merge-group runs failed the compiler differential on example/waffle-shop compile-store-warm stderr line 11 (CHI-570 flake shape, now at the store-warm step). Not reproducible locally (3/3 example runs pass with this change). Re-queue after #1069 lands so the next failure keeps its evidence.

[kvlonge]

Landed together in #1075 to cut merge-queue CI runs; review history stays here.