1/3 Regenerate the fixture corpus for MLPerf Endpoints v1.0

[arav-agarwal2]

Stack 1 of 3 · main ← 1-fixtures ← [2-checker] ← [3-docs]

Regenerates the 12 test_submissions/ trees for the v1.0 format, and commits the
script that did it.

Warning

CI is red on this PR, by design. The corpus is v1.0 while the checker reading it
is still v0.7 — 22 failures, all of that shape. It goes green in stack 2, which is
where the checker lands. Nothing merges to main until the tip is green.

How to review this

Read tests/tools/regenerate_fixtures.py and spot-check its output. Reading 9,000
lines of regenerated JSON and YAML is not a good use of anyone's time.

The script is committed and idempotent on purpose. The corpus gets hand-edited between
rounds, so a one-shot migration would be unrunnable the moment someone touched a
fixture, and the next format change would start from scratch. --check exits non-zero
if anything would change; a test in stack 2 asserts that, so idempotence is verified
rather than claimed.

What it does per tree

Per-point system description	Folds the per-system system_desc_id.json and each point's run_metadata.json into a per-point system_desc.json (policies PR #119), then deletes both sources.
§8.2 renames	system_availability_status → publication_status; flat node-level accelerator_* fields nest into accelerator_info[] (§8.2.1).
tps_utilization	Recomputed per Pareto curve from result_summary.json, so the stored value agrees with the measurement the checker recomputes it from.
§8.3 disclosure	Every point.yaml gains the ten §8.3 fields plus shared_src, shared_docs, seed_set, target_cohort.
§4.6 seeds	Rewritten under their v1.0 names from the mirrored seed set.
Metrics	TPOT migrated out of run_metadata into result_summary.json in nanoseconds; a TTFT P90 interpolated from the summary's own P50/P95.

Two things it cleans up that predate this migration

Stale artifacts. results_summary.json ×10 and point_<N>.yaml ×9, left over from
the pre-r<N> naming.

Missing src/ trees. Every sub_* fixture was failing src-dir — §2.2.1 has
required src/<impl>/README.md all along — for reasons no test asserted, so every test
had to look past an error unrelated to what it was checking. Each tree now has one.

One fixture needed repair, not regeneration

valid_standardized is the corpus's must-pass tree and stops passing under v1.0.
C_max=1000 with a derived C_min=16 puts Low Concurrency at 17–26, and its points
were 16/38/88/256/512/768/1000 — nothing lands in the window. v0.7's fixed 33–42
window caught 38, which is why it was built that way. Moving that point to r20 is the
smallest change that restores coverage, and it is recorded in POINT_RENAMES with that
reasoning rather than applied silently.

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅