fix(harness): classify beyond-engine primitives before capability_absent

[masterleopold]

Why

The test job on main fails 5 tests, all in the 4-way retarget gate: pour and stack-blocks (and the BEYOND fixture) are expected to classify as beyond_engine but come back as capability_rejected.

CI builds the rfl binding from the head of the public robotfoundationlayer/rfl repo (unpinned in .github/workflows/ci.yml). Upstream rfl now reports an unimplemented primitive as capability_absent: <primitive> when it is retargeted onto a descriptor that cannot declare it. In retarget_outcome, the capability_absent message check ran before the engine-coverage check, so a beyond-engine skill (place.put_down, place.stack) whose primitive is absent from the descriptor got tagged capability_rejected.

Per the documented gate order (CLAUDE.md gate #2), a valid emission the engine cannot lower is beyond_engine (a coverage gap, NOT a model error), and rfl / vendored descriptors are consumed-only. So the fix is entirely in cobel's classification logic.

What

Give the engine-coverage gate precedence over the embodiment-mismatch path:

except rfl.RetargetError as exc:
    msg = str(exc)
    # A skill using any primitive outside ENGINE_18 cannot be lowered by ANY
    # descriptor — coverage gap, not a real capability mismatch.
    if primitives_used(skill_yaml) - ENGINE_18:
        return Outcome("beyond_engine", msg)
    if "capability_absent" in msg:
        return Outcome("capability_rejected", msg)
    return Outcome("malformed", msg)

A skill built purely from ENGINE_18 still has an empty primitives_used - ENGINE_18, so genuine embodiment mismatches (seat-fuse, latch-buckle on a no-F/T arm) remain capability_rejected. sort-by-weight stays beyond_engine, and the malformed / retarget_ok paths are unaffected.

Verified by reproducing CI locally (rfl binding built from the public repo): full suite now 35 passed (was 5 failed, 30 passed).

---

^{Need help on this PR? Tag /codesmith with what you need. Autofix is enabled.}