[hongzuoj-pixel]

Summary

This PR implements the current P0 quality and evaluation work on top of upstream commit 3fca464.

• Add a lecture-note structure contract and deterministic document framing.
• Add guarded one-shot structure repair that rejects truncated or regressing output, preserves source markers and image targets, and re-runs image/figure grounding after an accepted repair.
• Add a fixed-sample evaluation harness, synthetic sample deck, human-review rubric, and Markdown/DOCX/PDF layout verification.
• Harden benchmark path handling and fail closed when required artifacts are missing or invalid.
• Clarify the boundary between coverage signals, heuristic quality checks, and manual review in the GUI and documentation.
• Expand CI to install GUI dependencies and run the complete test suite.

Safety and compatibility

• The branch is based directly on Cat-blizzard/SlideNote@3fca464; it does not include the stale history from the fork's main.
• Evaluation case IDs and output paths are validated before deletion or writing.
• Existing source markers, image references, and coverage evidence cannot regress during structure repair.
• Text line endings are normalized with .gitattributes to avoid repository-wide formatting noise.

Validation

• python -m pytest -q: 354 passed, 0 skipped
• python scripts/smoke_first_run.py: passed
• python -m pip check: no broken requirements
• GUI import smoke test: passed
• git diff --check: passed
• Secret-pattern scan: clean

Remaining real-world validation

The included synthetic/local benchmark verifies the engineering workflow, but it does not prove that LLM-generated notes are accurate for arbitrary real courseware. A production-quality claim still requires fixed-model before/after runs on several real lecture decks, manual fact/source/figure checks, and actual DOCX/PDF export review.