Restore the shape777 report schema the committed evidence and verifier read

[pxtroniwnl]

Summary

benchmarks/shape777.py and benchmarks/shape777_reranker.py no longer emit the
schema that the committed evidence uses. results/raw/shape777-direct.json and
results/raw/shape777-reranker.json were produced by an earlier, richer version
of these scripts, and in the meantime the generators lost fields, mode names and
top-level keys.

The important part is not that the JSON looks different. It is that
benchmarks/verify_published.py reads those raw files positionally and
re-derives all 69 headline claims from them, while the scripts that are supposed
to regenerate them had drifted. The published numbers could no longer be
reproduced from a run of the repository as it stands, and nothing detected it:
the verifier checks the committed bytes, never the generator.

This restores the contract the evidence was actually written against, and adds
the first test that ties the two together.

What changed

benchmarks/shape777.py

• report version shape777-direct-v1
• mode names restored to fresh_batch1, serial_prefix, parallel_suffix,
  matching the evidence
• COMPARISON_TOLERANCE / COMPARISON_NOTE, and the per-mode and top-level field
  sets that the verifier's positional reads expect
• judgments_per_second, p50/p95 latency, and the states field
• cache_hits for the suffix mode and true suffix token counts, so the
  throughput denominator is the number of real decoded tokens rather than a
  padded one

benchmarks/shape777_reranker.py

• report version shape777-reranker-v1
• timing_scope, states, max_tokens, yes_no_pair_forwards

Two judgement calls worth reviewing

passed means exact reproduction, not "within tolerance". The schema has a
passed flag per mode and a tolerance of 1.0, and the tolerance is explicitly
documented as non-gating. I therefore computed passed as bit-exact agreement
between compared cells and kept the tolerance as a recorded, non-gating field.
If the intent was for passed to be tolerance-based, that is a one-line change,
but it would change the meaning of a headline field and I did not want to make
that call silently.

mean_probability_difference is the mean over cells, not over tokens. I
verified this empirically against the committed evidence rather than guessing:
the per-cell means reproduce the published value, per-token weighting does not.

Verification

tests/test_benchmark_schema.py adds 6 tests that assert the emitted document
against the committed evidence. All 6 fail against the current scripts and pass
with the restored ones, which is the property that matters here: they encode the
contract rather than restating the implementation.

Also checked by hand: both scripts compile, --help works, the create-only
guard still refuses to overwrite an existing output path, and the
37 states x 21 options fixture validates.

pytest -q on this branch: 71 passed, 3 skipped.
python benchmarks/verify_published.py: 69 claims verified.
cd results/raw && sha256sum -c SHA256SUMS: 23 of 23 OK, zero failures.

Limitations

• I could not run either benchmark end to end. This machine has an RTX 3050
  with 6 GB of VRAM, which cannot hold Qwen3.5-4B in BF16, and I did not install
  torch because the CUDA wheel alone is 5-6 GB against 7.2 GB free. The
  restored code is therefore validated against the committed evidence and by
  static checks, not by a fresh run. Someone with adequate hardware should
  confirm the regenerated report matches the committed one before this merges.
• The per-row prediction schema is still not fully restored. I found that
  shape777-direct.predictions.jsonl also carries batch_size,
  padded_tokens, allowed_token_mass, answer_token_ids and
  full_vocab_argmax_id, and the current score_shared result does not emit
  them. I left that alone to keep this PR to the report contract, but the drift
  is not fully closed by this change and is worth a follow-up.
• The two source files and the new test conflict with open PRs llama.cpp backend: GPU offload and sequence-copy fan-out for shared mode #34, Pluggable direct prompt: built-in en/fr wordings, prompt files, one prompt_version per wording #35, Add Intel Arc (XPU) support to the Torch scoring paths #10 and
  Add Apple Metal (MPS) and CPU support with auto device selection #6, which also touch benchmarks/shape777.py. Expect a rebase.

No files under results/ are touched, no checksum changes, and no reported
number is altered by this PR.