xp_work behavior evals

Deterministic, single-turn behavior tests for the xp_work skill. Run via:

./scripts/run_behavior_evals.sh dev/xp_work

Each case lives in evals.json:

• id — stable case number
• name — human label
• prompt — what gets fed to claude -p
• fixture (optional) — directory under fixtures/ whose contents seed the test workspace
• assertions — list of structured checks against the post-run workspace

Assertion types: file_exists, file_absent, file_contains (literal substring, pattern field), file_not_contains, file_matches (multi-line regex, regex field).

Adding a case:

1. (Optional) create fixtures/<case-name>/ with the starting state of the test workspace (typically a work/ tree).
2. Append an entry to evals.json with a fresh id, a clear name, the prompt, the fixture reference, and the assertions.
3. Run ./scripts/run_behavior_evals.sh dev/xp_work and confirm it passes.

Multi-turn flows (e.g. pluck: with its planner-mode gate) are intentionally out of scope for this first cut.