feat(ocr): merge domain-adapted OCR upgrade

[PracticalSwan]

Summary

• freezes a deterministic 400-page public OCR benchmark and records bounded detector, general-recognizer, and Thai-recognizer trials
• keeps the original detector and general recognizer as the global default; exposes the accepted synthetic-only Thai recognizer only through explicit custom/adaptive profiles
• rebuilds public-only OCR-realistic LayoutXLM streams and selects the fresh four-epoch checkpoint
• records public-only calibration, one-time locked TEST evidence, 18-angle grids, a fixed 100-page unseen CORU run, and aggregate-only private operation
• updates CLI, GUI, portable-package, documentation, tests, provenance, privacy, and verification evidence

Accuracy boundary

This is an academic pre-model, not a production-ready document system. The locked 1,760-page image-to-JSON run completed without failures but measured polygon F1 0.3815, text coverage 0.1663, WER 0.9692, entity F1 0.0944, relation F1 0.0111, and canonical accuracy 0.2534. The requested quality targets were not reached.

Privacy and provenance

• public data only for fitting and selection
• Gmail fit rows: 0
• private operation remains bounded and aggregate-only
• no raw/private paths are tracked
• model, calibration, data, reports, and portable evidence are hash-bound
• K-Means remains display-only and does not control OCR or extraction

Verification

• complete information-extraction verifier: 81/81 passed on 2026-07-29
• host suite: 375 passed, 2 skipped
• changed/new OCR tests: 207 passed
• focused acceptance regressions: 56 passed
• execution evidence ledger: 30/30 passing records
• portable ZIP: 1,159,061,897 bytes
• portable SHA-256: d539c54f02c8e5bd204266eaed7e7372c4fd077d3cfa4062dccb1f894eb7d746

The locked TEST commands were not rerun during this pre-merge verification.