hyperloom custom workload: a result names its model by file name only, not by the fingerprint its own quality gate is keyed on #993

Class: the artifact does not describe the measurement (same as #892, #930, #932, #947, #954, #977 — all fixed).

The result JSON names the weights only through model_id, a hardcoded string:

$ python3 -c "import json;d=json.load(open('.../window-20261102/arm_pair2_c20-A.json'));print(d['model_id'], sorted(d['bench_summary']))"
Qwen3.8-27B-UD-Q4_K_XL.gguf (HRX0, gfx1151) ['levers','longctx','placement','pp512_tok_s','runner_type','score_geomean','serve_tok_s','tg128_tok_s']

No path, no hash, and the (HRX0, gfx1151) part is a literal in the source, not a measurement. Across every result
artifact this lane produced on this box: 140 artifacts carrying a bench_summary (150 result files scanned, from 70
distinct measurement directories), 0 with any model field.

Meanwhile the gate is keyed on the file's fingerprint: ref.json carries "model": "0bc73cb2357258a977f33f910a7d2b02c3c3ac7db98e4dcac6b58a8618b43827" and main() drops the stored references when the
measured file's sha256_head no longer matches it (if ref.get("model") not in (None, fingerprint) → ref = {}). So the
report says whether the references still apply, in the negative direction only (a WARNING line in a log), and the
positive claim ("this score and this KLD verdict belong to file X") exists nowhere in the artifact.

That is not hypothetical on this box: the 31 Oct restore took the weights and left ~/lb/hl-gate/.../{kld-base.kld, texts.json,ref.json} in place, so the weights were re-downloaded while the references survived — the same file name here
can legitimately be a different file, and every number of record (44.69 baseline, 50.87 best, the 18.8k long-context
pair) is tied to its weights by prose in a state file rather than by the artifact.

Wanted: every result names the model file it measured — resolved path plus the sha256_head fingerprint the gate uses, so
a result and a reference can be matched by hash instead of by file name.

Adjacent, measured while checking this, not fixed here: bench_summary.placement is null in 140 of 140 artifacts
— the placement counts only surface at the library's INFO level and 1bit serve has no -lv passthrough — so in practice
the artifact never states the HRX placement it was measured with, and a null cannot be told from "not captured".

Out of scope here: model_id itself is Hyperloom's own display string and is consumed by the report, so it is left as is.