PAIDEIA MCP v0.6.0

Highlights

• Dropped the codex-native OCR engine. It deferred OCR to a host's bundled vision (Codex CLI / ChatGPT subscription), which Alt's plugin sandbox can't provide — so as the default it produced empty ingest/grade output for the server's actual audience. OCR now always runs fully in-process.
• qwen3-vl is the new default, with automatic tesseract fallback. When Ollama is unreachable or the model isn't pulled, qwen3-vl degrades to tesseract so a default install still produces output. The two-engine set is self-sufficient.
• Honest engine reporting. run_ocr now returns the engine that actually ran; the converted-markdown provenance header and the grade confidence tier reflect a qwen3-vl → tesseract fallback (tier low) rather than mislabeling it.
• Legacy .course-meta OCR_ENGINE values (claude / openai-vision / codex-native / codex) now map onto qwen3-vl instead of the removed engine.
• Tool/grade schemas and init_course / bootstrap_alt_course defaults updated: engine enums are now ["qwen3-vl", "tesseract"], default qwen3-vl.
• README aligned with the OPTIMETA house format, plus a new README.ko.md Korean mirror; Engines section rewritten for the two in-process engines and the fallback.

Breaking changes

• codex-native is no longer a valid engine value. Passing it to ingest_pdfs / grade_pdf / init_course / bootstrap_alt_course raises; existing course folders pinned to a host-vision engine are transparently remapped to qwen3-vl on read.
• grade_pdf no longer emits the rasterize-only mode / page_paths; it always returns mode: "ocr-complete".

Verification

• `python3 -m pytest -q`: 44 passed
• `python3 scripts/smoke_stdio.py`: 54 tools, 5 prompts, 2 resources