Raw output backing the published figures. Every file written by the current
harness carries a provenance block — model path, quantization, thread
affinity, build SHAs, host, environment — so a number can be traced to the run
that produced it and invalidated later if the run turns out to be wrong.
Anything with "retracted": true names the reason and its replacement.
profile_w4a8_g128_p{1,2}.json are the clearest case: both show
_C::onednn_mm at ~2.97 ms/step — the signature of a checkpoint whose output
head was left in bf16 — and every W4A8 number derived from them is void. The
valid replacement is profile_w4a8_g128_head_p1.json.
They are kept because a deleted mistake teaches nothing, and because the older
reports still cite them.
