Reject invalid translated chunks before merge

[cursor]

Bug and impact

• Whitespace-only translated chunks were treated as existing outputs by selective planning and then silently skipped during merge.
• Severely truncated chunks only emitted warnings, allowing output.md and final book artifacts to be generated with missing or heavily incomplete content.

Root cause

• run_state.py only checked output file size, not trimmed text content.
• manifest.validate_for_merge() downgraded suspiciously short outputs to warnings and did not reject whitespace-only outputs.
• merge_markdown_files() continued after read/blank failures and could write a partial merged file.

Fix

• Reject unreadable and whitespace-only outputs in manifest validation, merge, and run-state recording/planning.
• Promote <10% output-size checks to hard errors for substantive source chunks while preserving warnings for tiny chunks.
• Add regression tests for blank outputs, stale output.md removal, legacy merge fallback, and suspiciously short manifest outputs.

Validation

• python3 -m unittest tests.test_run_state tests.test_merge_and_build -v
• python3 -m unittest discover -s tests -p 'test_*.py' -v (210 tests passed, 3 skipped)
• python3 -m compileall scripts tests
• Replayed the original blank-output repro: merge now fails, no output.md is written, and run-state schedules the blank chunk for translation.
• Ran the standard Alice mock full pipeline: merge/HTML/DOCX/PDF succeeded; EPUB failed due to the documented Ubuntu Calibre 7.6.0 packaging bug.