Split verifier support and hard-gate evidence rewards

[morluto]

Problem

The shared verifier validation modules had become large mixed-ownership hotspots, and mathematical benchmark rewards could still encode evidence as a soft additive term. That representation obscured the fail-closed invariant and made future verifier-policy migrations easier to get wrong.

Closes #500.
Closes #538.

Solution

Split the mathematical-benchmark and public-reproduction validation support into behavior-owned modules for fixtures, metadata, paths, verifier execution, evidence binding, input binding, assurance, and submission attacks.

Migrate the affected benchmark verifiers to hard-gate invalid evidence before reward aggregation. Remove redundant evidence terms from four remaining counterexample verifiers, strengthen the reward ratchet to recognize the ev alias, and synchronize the affected verifier checksum labels.

Testing

uv run --locked python -m pytest benchmarks/validation/test_fail_closed_reward_ratchet.py
uv run --locked python -m pytest \
  benchmarks/validation/mathematical_benchmarks_v1/test_subspace_direct_sum_counterexample.py \
  benchmarks/validation/mathematical_benchmarks_v1/test_polynomial_tail_counterexample.py \
  benchmarks/validation/mathematical_benchmarks_v1/test_matrix_square_zero_counterexample.py \
  benchmarks/validation/mathematical_benchmarks_v1/test_log_exponent_recovery.py

The ratchet reported 2 passing tests and the four affected verifier modules reported 5 passing tests. Earlier focused validation of the support split reported 1,699 mathematical-benchmark checks and 11 public-reproduction checks passing.

Trust & Compatibility Impact

This changes evaluation integrity behavior: invalid evidence now contributes zero reward rather than being represented inside a weighted aggregate. It does not authorize checkers or promote evaluator output to verified evidence. The support split preserves the existing public validation facade.

Checklist

• Routine local validation passes (make check)
• Relevant focused tests are listed above

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: aa6a565754

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".