fix: pro OCR ROI calibration and 24-word mnemonic pipeline fixes

[wabicai]

Summary

Device-calibrated OCR scenes for pro plus a chain of recognition/merge fixes for the 24-word create-wallet flow. Root causes were found by calibrating against saved captures and full-pipeline reproduction; final state: 12/18/24-word flows recognize all words correctly and the merged 24-word phrase validates against the BIP39 checksum.

OCR scene calibration (pro)

• Per-layout ROIs measured from real captures: mnemonic 12 (225,690 665x580), 18 (225,580 665x830), 24 part1 (225,695 665x715, stops above the 继续 button), 24 part2 (225,520 665x890, full list viewport), verify-number (160,530 520x170)
• Scene selection is word-count and part-aware (allowPartial picks part1/part2); pro2 behavior unchanged
• Calibrated scenes use base + horizontal-only expansion candidates (vertical expansion re-included noise the calibration deliberately excluded and blurred >1280px crops lexicon-snapped to wrong words, e.g. gym→game)

Recognition pipeline (paddleocr_en_infer.py)

• Blue-channel input for mnemonic layouts: screen text is cyan, warm-light glare lives in R/G — recovers words washed out by reflections (plain grayscale does not)
• Per-column row recognition: whole-row crops made the rec model glue the left word to the right index ("gym13. stove") which broke renderer index parsing
• Cross-row column-anchor filtering drops glare det boxes; viewport-clipped scroll rows are dropped (their half-glyphs parsed as phantom indexed words)
• scriptVersion in the server ready message, logged by the main process to make stale-server debugging obvious

Renderer fixes (CameraPanel)

• Wordlist decoupled from requireBip39: 24-word part1 (requireBip39: false) was treated as a SLIP39 flow and every valid BIP39 word got snapped onto the SLIP39 wordlist (skull→skunk, young→round…) — the root cause of the recurring wrong words at indices 1/2/13/14
• Isolated phantom-index guard before part merge: clipped scroll rows parsed as e.g. "1. ens" and overwrote the correct stored word 1
• Letter→digit seam re-splitting before index parsing; removed the dead 12-word column-row fallback; diagnostic logs for capture consensus and final merged validation

Sequences

• 500ms delayBefore PIN entry (both first entry and re-confirm); ControlPanel executor now honors delayBefore (previously only the MCP executor did)

Assets

• PP-OCRv6 small det/rec onnx models (runtime default) and BIP39/SLIP39 wordlists

Test plan

• Reproduced every historical failing capture through the full pipeline (12w glare, 18w row-9 clipping, 24w part1/part2 across four separate runs): all words correct
• 24-word merged phrase validates with @scure/bip39 + english wordlist
• MCP flow consistency: all sequence IDs referenced by hardware-js-sdk automation (/health sequenceIds drift check) still present
• tsc --noEmit clean; python py_compile clean

🤖 Generated with Claude Code

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	pypi/​onnxruntime@​1.27.0

View full report