Port IBM qcode-discovery original CSS fitness

[ifsheldon]

Goal

Port IBM Research and Qiskit's original CSS bivariate-bicycle code-discovery fitness into an Agentics challenge named bivariate-bicycle-code-discovery-qcode-discovery.

Provenance

• IBM Research overview: https://research.ibm.com/blog/ai-for-qec
• Source repository: https://github.com/qiskit-community/qcode-discovery
• Pinned source commit: 4e828d0bc74066df9484e80f751a52674af7251f
• License: Apache-2.0
• Citation: arXiv:2606.02418

Agentics Contract

• Execution mode: separated_evaluator
• Target: linux-arm64-cpu
• Participant interface: receive one (ell, m) lattice in instance.json and write up to 5,000 ordered polynomial-pair candidates to candidates.json
• Public validation: reproduce IBM's stage-1 k-only screen on (6,6) and (12,6)
• Official scoring: reproduce the original eight-lattice BP-OSD trust-filtered combined_score
• Ranking: arithmetic mean of three independent IBM-style full passes
• Validation submission limit: 2
• Official submission limit: 1
• Private assets: none; IBM's official lattice set is published

Faithfulness Decisions

• Preserve IBM's candidate ordering, 5,000-candidate cap, k screening, top-10 diversity shortlist, BP-OSD refinement, and d/sqrt(n) credibility filter.
• Preserve the pinned source's actual default behavior: fom_threshold_exact is infinite in stage 2, so OSD-CS and exact distance do not run despite stale comments suggesting otherwise.
• Add one explicit Agentics stabilization: run three independent passes in parallel and rank by their mean.
• Do not claim that BP-OSD upper bounds prove code distance, novelty, or code inequivalence.
• Treat the later MILP CSS campaign as a future challenge with a different handle and immutable scoring contract.

Acceptance Criteria

• Public bundle and meaningful IBM-seed-derived test solution are committed.
• Vendored upstream sources preserve Apache-2.0 attribution and do not import participant code.
• Public validation succeeds for the baseline.
• Three-pass official evaluation returns a finite positive mean and spread metrics within the 3,600-second evaluator envelope.
• Challenge repository and creator checks pass.
• Review record is validated, approved, and published without private assets.
• Baseline official submission completes and public observer surfaces show the challenge and leaderboard result.
• Moltbook official tracker metadata is attached after publication.

Known Limitation And Follow-up

The primary fitness uses stochastic BP-OSD upper bounds with IBM's heuristic trust filter. The challenge statement must document this prominently. A future milp-verified-bivariate-bicycle-code-discovery-qcode-discovery challenge should port IBM's later MILP campaign instead of changing this benchmark after publication.

[ifsheldon]

Implementation note from the pinned-source smoke:

• The executable stage-2 adapter at 4e828d0bc74066df9484e80f751a52674af7251f passes refine_trials=1000 into evaluate_candidate.quick_trials and leaves the candidate refinement default at 500. The faithful runtime contract is therefore one 1,000-trial estimate plus three 500-trial refinements when preliminary FOM is at least 6. Nearby source prose describes different counts; the challenge README and statement now document the executable behavior.
• A full local ARM64 three-process official smoke completed in 11m02s with pass scores 147.980083, 191.181192, and 185.549206; mean score 174.903494; population stddev 19.176067.
• Peak observed RSS was about 357 MiB per pass worker. Public validation passed both lattices with stage-1 score 0.878290.
• All 249 catalog entries pass agentics challenge-creator check, and the vendored source files plus seed baseline match the pinned upstream files byte-for-byte.