fix(metrics): slice to n on the cache-hit path in the MMLU/GSM8K loaders

[Nitjsefnie]

Closes #14 — the follow-up you asked for in #13's merge note:

yes please, I'd take the follow-up that slices to n on the cache-hit path (with a too-many-rows fixture)

The change

Both loaders returned every row in the cache file on a cache hit, never checking the count against n:

-        rows = json.loads(path.read_text())
+        rows = json.loads(path.read_text())[:n]

One line each in load_mmlu_slice and load_gsm8k_slice. Since n is part of the cache key this is invisible in normal use — but a stale or hand-edited cache file made load_*_slice(n=1) return however many rows the file held, silently violating the contract the accept-path tests from #13 assert.

Tests — the too-many-rows fixture you specified

• test_load_mmlu_slice_cache_hit_slices_to_n
• test_load_gsm8k_slice_cache_hit_slices_to_n

Each seeds a cache file with 2 rows, calls the loader with n=1, and asserts exactly one comes back. They reuse the existing _block_network helper and the same tmp cache_dir pattern as the #13 accept-path tests, and sit directly after them in the same section.

They bite: against main's metrics.py with these tests in place, both fail with assert 2 == 1; with the fix, both pass. (Verified by checking the production file out from upstream/main, running, then restoring — not just by watching them go green.)

Gates

ruff format --check . → 0 · ruff check src tests → 0 · mypy src → 0 · pytest → 0 (58 passed, 2 skipped — your 56/2 baseline plus these two). uv.lock churn from uv run was reverted and is not in the diff. Diff is 2 files, +46/−2.

---

Implemented with AI assistance (Claude Sonnet 4.5), and independently mutation-verified by me against upstream/main before submitting.

[bamdadd]

Reproduced the mutation check locally — reverting the [:n] slice makes both new too-many-rows tests fail with 2 != 1, so they pin the contract rather than decorate it. The normal case is untouched ([:n] is a no-op when the file holds exactly n rows, and still returns everything when it holds fewer, matching the download path's min(n, len)). ruff/mypy/pytest all green here. Clean fix for the gap you spotted in #13 — merging. Thanks Peter.