v0.6.66b483

What's Changed

• ci: auto-dispatch publish + fan-out from release-candidate by @tobocop2 in #281
• README: trim implementation detail and deduplicate by @tobocop2 in #282
• Surface 'model not installed' instead of generic Internal error in chat stream by @tobocop2 in #283
• Classify provider errors into kinds, with clear user-facing messages by @tobocop2 in #286
• Bump llama-cpp-python to 0.3.23 by @tobocop2 in #288
• Fix Windows crash in GGUF architecture probe by @tobocop2 in #290
• Fix chat auto-scroll stalling mid-stream by @tobocop2 in #294
• crawler: clean reference links + surface first-run browser setup by @tobocop2 in #293
• Correctness & security audit of the non-multi-GPU paths by @tobocop2 in #289
• Enable flash attention for vision OCR by @tobocop2 in #295

Full Changelog: v0.6.66b481...v0.6.66b483