Orphan fixtures and minor scoring smells

[rendrag-git]

Cleanup grab-bag

Three small items found during the M3 inventory audit. None affect runtime correctness; all are convention drift / dead code.

1. `fixtures/needle_repo_8k` is unreferenced by any tier manifest

Only `manifests/openclaw-agent-core.json` (the legacy non-tier suite) uses it. Tier-medium runs needle at 16k and 32k; tier-small runs needle at 4k. There is no 8k slot. Either:

• Add an 8k task to tier-small as the smaller stretch case, or
• Delete the fixture as dead seed data.

2. `fixtures/real_repos/kingshot-ams-snapshot` is unreferenced by any tier manifest

Only `manifests/real-repo-readonly.example.json` (a template) uses it. GOAL.md M6 calls for "real-repo coverage beyond one snapshot" — this is M6 territory and tracked there. Keeping the fixture for now, but flag it once that milestone activates.

3. `_failure_type` has a duplicate `"preserved file"` branch in scoring

`openclaw_bench/scoring.py` `_failure_type` (around line 651) checks `"preserved file"` substring in two branches. The first match always wins, so the action_gate-specific second branch is unreachable. Same outcome (`instruction_violation`) either way, so no scoring impact, but the dead branch should be collapsed.

4. AGENTS.md "intent" comment convention not fully covered

AGENTS.md says every scoring branch should have a one-line "this catches X failure mode" comment. The branches at `scoring.py:40–46` (hallucinated, timeout, error) and the catch-all `else` at L112 lack the comment. Three lines of explanatory text would close the gap.

Severity

P3 — none of these affect any runtime behavior. Nice-to-haves before the next major refactor of scoring.py.

[rendrag-git]

Codex triage: category=enhancement (maintenance/cleanup). Impact=P3 (no runtime correctness impact; reduces dead weight + scoring confusion). Suspected areas: fixtures/needle_repo_8k, fixtures/real_repos/kingshot-ams-snapshot, manifests/*.json references, openclaw_bench/scoring.py (_failure_type duplicate 'preserved file' branch), and AGENTS.md scoring-branch intent comment convention. Next action: (1) decide keep vs delete orphan fixtures (or add an 8k task), (2) collapse the unreachable duplicate branch in _failure_type, (3) add the missing intent comments to the few scoring branches called out.