fix: honour the canonical-pin grace window in eval.verify (honest training PRs are auto-closed as eval:REJECT)

[kai392]

SparkDistill contribution

Track

Select the one track this PR belongs to.

• Dataset track submission
• Training/evaluation improvement

Neither box applies: this is a bug fix to the training-track gate itself, not a
submission. Per CONTRIBUTING.md this is a non-quality PR and scores 0 — it is
filed because the bug auto-closes other people's honest submissions.

Summary

Closes #231. Restores the canonical-pin grace window from #121 (issue #118) on the
one code path that still ignored it.

Root cause

#121 added a merge-base→HEAD grace window so a training-track PR is not rejected
when a dataset-track merge advances datasets/canonical.json mid-training. The
window reaches two of the three mix_manifest.sft_sha256 checks:

check	window applied
validate_pr_body_canonical_pin	✅
verify_remote_proof_bundle	✅
check_canonical_dataset_claim inside eval.verify.verify_submission	❌

gate_training_pr computes acceptable_sft_shas, but verify_remote_proof_bundle_scores
never received it, so _download_and_verify_bundle called
verify_submission(bundle_dir, frontier, attestation=attestation) — and verify_submission
had no such parameter. check_canonical_dataset_claim then fell back to
allowed = {canonical_sft_sha256()}: the single pin at the checked-out tree, which
training_track.yml checks out from pull_request.base.ref (live main).

Reproduction with a bundle pinned to the merge-base sha and a window of {base, head}:

stage 1  verify_remote_proof_bundle       -> []            # accepted
stage 2  verify_remote_proof_bundle_scores -> eval:REJECT
         'eval.verify canonical_dataset_failed: bundle mix_manifest.sft_sha256 does not
          match an accepted canonical pin (allowed 1 pin(s) for this PR window)'

The error text reports allowed 1 pin(s) — proof the grace set never arrived.

eval:REJECT is in _AUTO_CLOSE_EVAL_LABELS, and training_track.yml runs the gate
with --close-on-reject, so the honest PR is closed automatically.

Fix approach

Thread the already-computed window down the call chain, nothing more:

gate_training_pr → verify_remote_proof_bundle_scores → _download_and_verify_bundle
→ verify_submission → check_canonical_dataset_claim (which already accepts
acceptable_sft_shas).

record_merged_ledger_entry shares _download_and_verify_bundle, so it gains a
merge_base_ref argument and computes the same window via the existing
_canonical_sft_sha256s_for_pr_window helper; training_track_ledger.yml passes
pull_request.base.sha. Without this, the pin can advance again between the gate job
and the ledger job, recording eval:REJECT for an already-merged run and skipping its
runs/frontiers.json update — which would then mis-tier every later submission.

Impact

• Honest training-track PRs are no longer auto-closed when a dataset-track PR merges
  during their train→eval cycle. This is what the PR template already promises:
  "any pin from merge-base through HEAD is accepted".
• Miner credibility (merged/(merged+closed), which gates eligibility at 80% and
  multiplies earned score) is no longer damaged by an automated wrong-close.
• The merged ledger and frontiers.json stay correct for runs that were merge-eligible
  at gate time.

Risk and tradeoffs

• Backward compatible. acceptable_sft_shas and merge_base_ref both default to
  None, which preserves today's strict behaviour. The standalone
  python -m eval.verify validator path is unchanged and still requires the pin at
  datasets/canonical.json.
• No security loosening. The accepted set is bounded by the PR's own
  merge-base→HEAD history of datasets/canonical.json; a miner cannot inject an
  arbitrary sha, and any pin outside the window is still canonical_dataset_failed
  (covered by a test). The gate and the verify stage now agree on one set instead of
  silently disagreeing.
• No new dependencies; no behaviour change for dataset-track PRs.

Tests

Three regression tests in tests/test_training_track_pin_grace.py (the file #121 added):
verify_submission accepts a merge-base pin, still rejects a pin outside the window,
and verify_remote_proof_bundle_scores no longer tiers a base-pin bundle eval:REJECT.
All three fail on main and pass with this change.

pytest tests/ -q       269 passed, 8 skipped, 1 pre-existing env failure
                       (test_aggregate_and_publish_mining_dataset needs a SparkProof checkout)
pyright eval/verify.py eval/training_track_gate.py eval/record_training_ledger.py
                       0 errors, 0 warnings

One existing test fake (fake_verify_submission) gained **kwargs to accept the new
keyword argument.