Fix syzygy verifier handoff metadata

[yuelgrace1810-ops]

What changed

• correct the graded Jacobian-syzygy result metadata to say that verification consumes the exact producer input and inline candidate, rather than a nonexistent result_uri;
• update the independent verifier description to state the same contract;
• add a regression covering both the producer hint and the installed verifier description.

Root cause and observed impact

A held-out evaluation of Dimca and Pokora's 2026 nine-line counterexample (arXiv:2607.01985) successfully discovered and ran polynomial.jacobian_syzygy.minimum_degree.compute for both arrangements. The model then inspected the independent verifier, but declined to call it because:

• the computed result advertised verification_input_field: "result_uri";
• the compute operation produced no result artifact URI because this is an ordinary inline computed operation; and
• the verifier description said it consumed a "stored producer result".

The installed verifier actually accepts the original producer input and its inline candidate. The conflicting metadata converted two available independent replay paths into COMPUTED-only final claims.

Scope and safety

This changes only agent-facing handoff metadata and the result's self-description. It does not alter the polynomial computation, candidate schema, independent checker implementation, authorization, assurance semantics, or verification record behavior. The checker still reconstructs every bounded coefficient map independently and remains fail-closed.

Validation

Focused regression:

• graded Jacobian-syzygy compute/verify test: 1 passed.

Planner-selected validation:

• unit: 868 passed;
• component: 759 passed, 3 expected platform skips;
• domain: 185 passed;
• composition: 475 passed, 2 expected Lean skips;
• storage: 125 passed;
• process: 234 passed, 1 expected platform skip;
• MCP: 46 passed;
• end-to-end: 8 passed, 1 expected Lean skip;
• Ruff, formatting, complexity, dependency, dead-code, mypy, test architecture, product architecture, package build: passed.

Overlap

Issue #883 tracks the wider audit of historical artifact-backed verification boundaries. This PR does not perform that audit or alter transport architecture; it fixes one directly observed contradictory inline handoff.

[yuelgrace1810-ops]

Independent held-out reproduction from a different source/domain instance: Abe–Denham's 2026 G(3,3,3) Ziegler-criticality counterexample, supplied as the exact rational polynomial (x^3-y^3)(y^3-z^3)(x^3-z^3) to gpt-5.4-mini (low reasoning).

On current main 614fdc40, the producer completed in ~0.45 s, but the arm followed the advertised verification_input_field: "result_uri" path, diverted to polynomial.jacobian_syzygy.coefficients.materialize, and then made three invalid verifier attempts roughly one minute apart. The frozen 240 s arm timed out. A separate 420 s diagnostic completed after 184.03 s / 12 MCP calls / 449,203 input tokens, still without verification, and explicitly reported that it could not satisfy the verifier's artifact shape.

This corroborates the PR's root cause and shows material efficiency/reliability impact beyond the original Dimca–Pokora arrangements. The exact producer/verifier computation itself was fast; the contradictory handoff metadata caused the loop. No additional code change appears necessary beyond this PR.

[yuelgrace1810-ops]

Refreshed this draft onto current main after a #929-only weak-model regression reproduced the remaining contradiction: the producer returned the correct degree-0 syzygy but still advertised verification_input_field: "result_uri", leading to an invalid verifier call and a 600 s timeout. Current refreshed head validation: focused graded-syzygy tests 2 passed; full MCP lane 47 passed. GitHub should now be able to evaluate the branch against current main.

[yuelgrace1810-ops]

Follow-up weak-model replay with both #917 and #929 present isolated one remaining handoff ambiguity. The model used the correct outer {payload:{input,candidate}} envelope, but manually selected only 8–11 fields from the 15-field producer output.result; it omitted required variables, searched_through_degree, status, and/or source_kind. Three verifier calls failed at candidate, and the run ended COMPUTED after 424.94 s / 11 MCP calls / 588,385 input tokens (534,528 cached). I amended this PR so the producer now advertises verification_input_field: "input_and_complete_output_result", and the verifier description explicitly requires the complete, unmodified producer output.result object. Focused compute plus actual verifier replay: 2 passed. The earlier refreshed full MCP lane remains 47 passed; this amendment is metadata/test-only.

[yuelgrace1810-ops]

Post-amendment combined replay with #929 confirms the handoff fix: the weak model passed the complete producer output.result to the verifier on its first attempt, which returned checker-authorized VERIFIED. Runtime fell from 424.94 s / 11 MCP calls / 588,385 input tokens to 90.08 s / 8 calls / 210,174 input tokens. Important scope caveat: this trajectory is transport evidence only, not a valid mathematical evaluation—the model misbound the polynomial as a binary form (d/dz = 0, kernel (0,0,1)) instead of the required X=x+y, Z=y+z lift (kernel (1,-1,1)). The checker correctly verified the exact different input it received. That separate prompt-to-call binding failure is not caused by this PR and is not being counted as conjecture success.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e5cf77042e

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".