[kvlonge]

Why

The compiler differential intermittently reports a warm-compile stderr difference that we cannot reproduce locally or in diagnostic CI runs. The next natural failure should carry its own evidence instead of one differing line.

Changes

• --evidence-dir PATH: when a project differs, copy each engine's target/ (including the SQL-test artifact cache), every command's stderr, and a compiled-tests-stat.tsv listing size, mtime_ns and ctime_ns of every compiled test file.
• The differential job passes it to both differential steps and uploads the directory as an artifact on failure only, following the existing upload step.
• No comparison logic changes: evidence is written after differences are computed and never affects the result.

Verification

• New parametrised harness test: a differing project keeps both engines' evidence (target files, stderr, stat listing); identical runs keep nothing.
• Differential script unit tests (415 passed); a real --corpus examples --evidence-dir run passes and writes nothing.
• make check-ci; hygiene scan.
• No test expectations or thresholds changed.