Establish versioned verification framework

[jordanyushen]

Summary

Establishes the first versioned verification framework without changing solver equations or engineering behavior.

• adds strict JSON schemas for canonical-SI case inputs and expected quantities
• adds typed verification models, loader, comparator, runner, and python -m beamcheck.verification
• migrates the four existing classic regression inputs as explicitly draft/unreferenced corpus assets
• rejects NaN/Infinity, invalid schemas, empty verified expectations, owner-reference placeholders, and unreviewed reference IDs
• runs the verification corpus separately in Windows CI

Files changed

• beamcheck/verification/ — loader, models, comparison engine, runner, module entry point
• verification/ — schemas, four draft cases, reference placeholders, generated-summary policy
• tests/test_verification_corpus.py — corpus invariants and pass/fail runner behavior
• .github/workflows/tests.yml — verification CI step
• requirements.txt — JSON Schema validation dependency

Engineering behavior changed

No. Production solver equations, validation, UI, diagrams, formula trace, and PDF calculation behavior are unchanged. All verification execution calls the existing production solve() function.

Tests

Commands:

• python -m pytest -q
• python -m beamcheck.verification
• python app.py --smoke-test
• python -m pip check

Results:

• 30 tests passed
• verification exit 0
• app smoke exit 0
• no broken requirements

Verification corpus

• Verified: 0
• Draft: 4 (BC-0001 through BC-0004)
• Failed: 0

Draft cases are intentionally skipped and are not advertised as independently verified.

Risks / unresolved

The corpus has no owner-approved independent references yet. Expected files remain empty by design, and no generated verification statistics are published in the README.

Owner action required

Provide or approve an independent source for each of the four classic beam cases, including title, author/organization, edition/revision, page/equation/section, and URL or identifier where applicable. Only then should a separate PR add expected quantities/tolerances and promote cases to verified.