hyperloom custom workload: a DFlash arm names no drafter, so the 50.87 best of record cannot say which of this box's two drafter quants produced it #977

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

HL_DFLASH is not a configuration on its own: serve_lever_args() hands --dflash the file HL_DFLASH_GGUF names,
defaulting to the relative path dflash.gguf (so the run's working directory decides it too), and HL_MTP_HEAD
substitutes an external MTP head for the model's own. levers() — the only place a result echoes its configuration —
lists HL_DFLASH and HL_MTP_MAX but neither HL_DFLASH_GGUF nor HL_MTP_HEAD.

This model has two drafter quants on this box (~/models/qwen38/DFlash2/{q8_0,q4k}.gguf, 2,056,414,560 B and
1,207,231,840 B), so the gap is not academic — and the tuner itself treats the pair as one configuration: it drops a
drafter-quant variant because overwriting HL_DFLASH_GGUF "would make its number incomparable to the baseline"
(hl-workspace-26030/... orchestrator log, 2026-10-30 18:14:39).

The arm of record, 50.87 (HL_DFLASH=1 HL_MTP_MAX=6), is exactly this shape. Its artifact:

$ python3 -c "import json;d=json.load(open('.../benchmark_custom_20261101_003158/inferencex_result.json'));print(sorted(d['bench_summary']),d['bench_summary']['levers'])"
['levers','placement','pp512_tok_s','runner_type','score_geomean','serve_tok_s','tg128_tok_s'] {'HL_MTP_MAX': '6', 'HL_DFLASH': '1'}

So which drafter produced 50.87 is recoverable only from serve.log line 29 of that session
(loading draft model '/home/bcloud/models/qwen38/DFlash2/Qwen3.8-27B-DFlash2-q8_0.gguf') — or from the session's
config.yaml, which is Hyperloom's file, not the measurement's. Until #892 a second arm in the same RESULT_DIR
overwrote that serve.log; 50 HL_DFLASH-carrying artifacts (25 distinct runs x benchmark_report.json + inferencex_result.json) exist in that workspace and none of them names its drafter.

Same class, smaller: the pp512/tg128 medians are over -r HL_REPS repetitions (default 3) and the result never says how
many, so two numbers a couple of percent apart cannot be told apart from a rep-count change.

Wanted: every result names the drafting files it really used (resolved path + hash, and whether the path came from the
environment or the default) and the repetition count behind its medians.