Fix numeric Run integrity and add optional mechanics consumer

[aliengineering-byte]

Scope

Fix a reproduced finite-numeric durable Run integrity defect and deliver a separate optional external mechanics consumer. GaugeMesh remains a domain-independent capability/model router. There are no domain formulas or new dependencies in the binary, no new architecture/database schema, and no changes to other products.

Evidence and correction

The actual public 0.4.0 Linux x64 binary accepted four consumer steps, then export returned GM_RUN_INTEGRITY. A generic finite-number regression reproduced the same error on SQLite reload. Enable the existing pinned serde_json dependency's float_roundtrip parser and retain the storage/repeated-export regression. This is a substantive 0.4.1 patch, not a metadata-only release. Immutable 0.4.0 artifacts remain unchanged; previously recorded neutral results are not relabelled as numerical success.

Optional consumer

Python standard library only, public MCP Tasks/Run interface, immutable four-stage plan, explicit units/assumptions/consent, original educational section model, independently derived Decimal active-set solution, audited nonconvergence, fully checked synthetic results, offline HTML/Markdown and structured evidence. It supports a real input checkpoint/gateway stop/resume without a duplicate input step. The gateway does not become a structural application.

Official ACI/ASCE catalog metadata only. No normative content, authorized standards rules, jurisdictional compliance or construction approval. Rule requests remain SOURCE_AUTHORIZATION_REQUIRED; exclusions stay NOT_EVALUATED; real decisions require qualified review. No restricted/private assets, credentials or user brief are included.

Verification status at opening

• Ten Python tests passed, including independent units, analytic fixtures, steel states, rounding, forged quantities, every synthetic-result field and nonconvergence trace mutations.
• Independent read-only review found no remaining blocker in that source scope; its Python execution passed. It did not execute Rust or claim public 0.4.1 delivery.
• The pre-fix numerical storage test is a recorded RED result. Patched workspace tests/build and source consumer acceptance are in progress; no post-fix/public success is claimed yet.
• Required CI now runs the optional consumer alongside the neutral external suites. Release qualification will use the actual fresh downloaded, attested public archive.
• Browser policy prevented local HTML visual inspection; only automated report checks are claimed.

Merge only through the unchanged normal required gates. Historical ResiliReplay remains 3 PASS / 10 FAIL / 13, PARTIAL. No crates.io work; both crates remain non-publishable.

[aliengineering-byte]

Source qualification update (not public-release proof):

• Patched workspace tests pass: 43 application tests, 5 transport tests, 92 core tests, no failures, one intentional helper ignored; doc tests pass. The finite-numeric regression that previously failed now passes.
• Eleven Python tests pass, including independent analytical/unit/material-state checks and full synthetic/refusal/report verification.
• A clean CI archive, version/source checked as 0.4.1/eaca468762ada89874817f7682229015f5d7626a, executable SHA256 a0b2a89cbe05eefb0261e9d5436fbd127105b69751201eb157d00d7046ef46f9, passed the full updated external consumer suite. The later report-only commit changes no Rust binary source.
• Four-step Run: 4 starts / 4 effects; independent calculation VERIFIED. Actual checkpoint interruption/resume: same Run, one input start, accepted input bytes unchanged, four total starts. Completed resume and offline verification added no worker events.
• Nonconvergence: Run verified its declared artifact outcome, numerical verification NOT_EVALUATED, HUMAN_REVIEW_REQUIRED. Five pre-execution refusals created no workers. Artifact, numerical report, verification-command, HTML and frozen-input tampering were rejected, including rehashed report manifests.
• Initial CI found a stale package-version assertion and nondeterministic report ordering after sorted-JSON reload. Both are corrected, with a deterministic rendering regression; original failures remain recorded. No integrity check or publication restriction was weakened. Only superseded running checks were cancelled; all current-head required checks must pass.

Independent review is source/Python scope, not a structural certification or public-release test. Normal merge, immutable release, fresh public download, attestation/checksum/source verification and external rerun remain outstanding. Standards rules stay SOURCE_AUTHORIZATION_REQUIRED.