Harden numerical helpers, add circulation contracts and streamline locked CI

[docxology]

SPM helpers reject malformed formulas and unknown interactions, produce finite degenerate bases, and terminate for the default polynomial expansion. MATH adds finite skew parts, Markov currents and discrete graph Hodge projections. Independent adversarial review found scale-dependent and near-dependent face-boundary failures; validation and projection now use the same numerical SVD space, with regression witnesses. H3/RNG guidance, executable examples, changelog, generated README/AGENTS inventories and readiness records match the implemented contracts.

CI shares locked CPU setup and per-interpreter package caches with one writer per key. Only writers prune; workflow tests reuse parsed YAML through isolated copies. Manuscript verification stops admission after interruption and retains the attempt. Real-child tests ensure ordinary exit code 5 remains failure, including allow_empty; pytest empty-selection policy remains limited to actual pytest entry points.

Final clean candidate: 42d2964b4a3919f3865b2b41a69acbe100b97c7c, preserving accepted GNN 4.1.0 and published ancestry.

Validation:

• All 21 hosted CI jobs, all 29 PR checks and separate whole-tree formatting passed.
• Local full-eleven producer passed: 45 unit suites / 10,870 cases, 1,249 integration cases, 54 performance cases and H3. ROOT manuscript 144 cases and native PDF 7 cases passed; producer drift check passed.
• All 46 coverage dispositions measured; all 45 unchanged module floors met. TEST raw coverage is 2,976/3,383 = 87.9693% in the candidate run and 2,977/3,383 = 87.9988% in the main run, both above the unchanged 85% floor. ROOT measured separately without an invented floor.
• All 23 main-CI raw archives match service digests; bounded ZIP safety, canonical receipt/report/log hashes and tested-tree identity verified. Both installed-wheel inventories verify 45 wheels, 45 base imports and 20 operation profiles.
• Supplementary Windows, portability and interchange archives audited. Windows verifies 52 positive cases per interpreter and intended optional-boundary failure separately. Hosted producer tier is default-seven; local full-eleven is separate.
• Native independent review resolved all identified source/infrastructure blockers, reran 35 analytical cases and 40 randomized cycle witnesses. Independent source/output snapshot audit verified 2,038 files / 133,886,502 bytes, source hash 7d76fd470a2bec35, 11 receipts, four figure hashes and 12 resolved manuscript files.

Published to main by ordinary fast-forward at the exact candidate SHA. All 21 same-SHA main-triggered CI jobs and separate Windows, portability, formatting and paired workflows passed; final main raw artifacts, coverage and producer identities are audited. Unmet issues stay open as requested. Implements GEO-INFER anchors for #58; external formal bridge edits and proof acceptance remain separate. Hardware, licensed data, configured-advisor access, historical causality, scientific review of every existing output, release tagging/publication and deployment retain their distinct requirements.