1.2.2

Cloud AI recognition no longer fails silently when a model runs out of output
budget — which reasoning models, increasingly common on OpenAI-compatible
services, hit routinely.

Fixed

• Cloud AI recognition now detects a response that was cut off by the output
  token limit and says so, instead of silently writing a half-finished
  transcription into the note (or reporting the generic "returned no
  transcription"). Reasoning models made this reachable: their thinking tokens
  count against the same budget. (#10)

Changed

• The per-page output budget for Cloud AI recognition is now 8192 tokens, up
  from 2048. It is a ceiling rather than an allocation, so unused headroom
  costs nothing; the old value left little room for models that reason before
  answering.
• SELF_HOSTING.md documents using the Custom endpoint vendor with a hosted
  OpenAI-compatible service, with Hetzner Inference as a worked example —
  including the exact-case model name, whose absence returns a 403 that looks
  like an API-key failure.