feat: add verified industry reference examples

[svonava]

Summary

• add runnable financial filing, prior authorization, and maintenance triage examples
• include checked evidence from real component calls against public SIE main at 4b5ea9c
• preserve each source, parsed output, raw model response, mapped record, review, manifest, and evaluation
• define narrow decision boundaries for financial, healthcare, and industrial evidence

Evidence and provenance

• Finance reads original values from the original Form 10-Q, reads restated values from the Form 10-K/A, and cross-checks the amended filing's previously reported column.
• CMS keeps each submitted document as a distinct source fragment and reproduces the published one-month documentation gap without making a coverage or medical decision.
• Maintenance reconstructs the NTSB detector sequence without adding a causal claim or control action. A saved structured-output probe includes its exact request, response, input hash, runtime commit, and narrow interpretation.
• Public configs use docling-project/docling. Manifests show the internal docling ID used by the direct local server during verification.

Verification

Live NVIDIA L4 runs:

• Finance: 8/8 evaluation checks, 2026-07-25T02:28:01.848776+00:00
• CMS: 10/10 evaluation checks, 2026-07-25T02:35:39.667778+00:00
• Maintenance: 8/8 evaluation checks, 2026-07-25T02:35:45.228957+00:00

Local package checks:

• Finance: 12 tests
• CMS: 18 tests
• Maintenance: 12 tests
• Ruff format and lint pass in all three packages
• every manifest artifact hash and diagnostic-fixture hash matches
• adversarial review found no remaining blocker, secret, unrelated file, stale artifact name, or chronological run path

[svonava]

@coderabbitai review