test(metadata): stop conflating drift with the warning gate

[mosheabr]

Problem

The --no-ai case in test_generate_skill_metadata.py asserted on main()'s bare exit code:

rc_no_ai = exit_code_for(["--check", "--no-ai"], None)
assert rc_no_ai == 0

--check returns 1 for output drift as well as for skill warnings. The behaviour being pinned is the warning gate — unenriched skills must not fail PR CI when no inference key is available — but the assertion cannot tell the two causes apart.

So any PR that edits skill content fails this test for reasons unrelated to enrichment. #445 is the first to hit it: 70 changed files including SKILL.md at 62+/147-, which legitimately drifts the checked-in metadata. My defect, currently blocking someone else's PR.

The irony is that the test was added in 276b9bc to prevent exactly this class of problem — a change that red-lines every open PR — and its own commit message says "adds test cases pinning both halves so neither can regress."

Fix

Neutralise diff_text for the duration of the run so the exit code is attributable to the warning gate alone. Drift is already covered by case 4.

What I got wrong first

The first attempt excused drift instead of isolating it:

r.check(..., rc_no_ai == 0 or drifted)

That was worse than the original bug. Mutation testing showed why: removing the --no-ai gate entirely — the exact regression the case exists to catch — still passed, because drift was present and the assertion excused it. A test that cannot fail is worse than no test.

Verification

Scenario	Expected	Result
#445's content, without this change	reproduce the failure	FAIL
#445's content, with this change	unblocked	PASS, rc=0
#445's content + --no-ai gate removed	still has teeth	FAIL
Clean main, with this change	no regression	PASS
Clean main + --no-ai gate removed	teeth on a clean tree too	FAIL
generate-skill-metadata.py --check --no-ai	generator unaffected	rc=0

Note on the with-AI case

Added a comment recording that unenrichable new skill with AI available -> fails the run asserts the outcome, not the route. In practice rc=1 arrives there via schema validation before the warning gate is consulted, so it does not pin the gate — the --no-ai case is what does. Pre-existing, documented rather than redesigned.

Scope

Test file only. The generator is untouched.

This does not guarantee #445 turns green — that PR still drifts metadata.json legitimately, so check may fail for a correct reason. This only removes my defect from the picture.