hyperloom custom workload: the quality-gate verdict does not name the reference it was measured against, so a KLD number cannot be compared across sessions #954

Class: the artifact does not describe the measurement (same as #892, #930, #932, #947 — all fixed).

benchmark.py decides pass/fail by comparing a candidate against two references kept in HL_STATE
(~/lb/hl-gate/<model-stem>): kld-base.kld (1.0 GB of base logits) and texts.json (the baseline's three greedy
replies). The result JSON records the candidate's own numbers only:

$ python3 -c "import json; print(sorted(json.load(open('arm_pair2_c20-A.json'))['quality_gate']))"
['kld_max_mean', 'kld_mean', 'kld_ok', 'kld_recorded', 'passed', 'reason', 'texts_differing', 'texts_identical']

So a row that reads KLD 0.000e+00 <= 0.0001, texts identical cannot be tied to the reference it is a distance to.
That state is per box, outside the checkout, and kld() re-records the base whenever kld-base.kld is absent
(recording = not base.is_file()), so the same KLD number can be PASS against one base and meaningless against
another — while the reports of the two series look identical. ref.json does carry a model fingerprint and the
baseline's levers, but neither reaches the result.

Wanted: every compared result names the reference it was compared against.