Fix HTML assessment validity, optimizer consistency, and session safety

[shepherd70]

The HTML tool could report PASS for malformed inputs or unsupported velocity claims, recommend designs that failed intake checks, and partially overwrite a session when loading invalid JSON. This implements the fixes from the UI review in #12.

• Introduce explicit assessment states, strict shared validation, and velocity evidence requirements. Fresh examples remain unassessed until confirmed; screen and print use the same verdict rules.
• Evaluate optimizer candidates against every intake check, preserve fixed/displayed dimensions, and preview changes before Apply.
• Validate saved sessions before replacing live state; persist checklist answers; align CSV parsing with Python using a shared fixture. Convert flow units, identify product snapshots, and provide deletion undo.
• Add input labels, field error associations, live status messages, focus preservation, responsive layouts, consistent decimals, and improved PDF pagination.
• Add Node and Playwright regression coverage and CI jobs; keep the distributed HTML self-contained and offline.

Validation: 46 Python tests, 34 JavaScript regression cases, and 13 Chromium browser workflows passed locally; ruff and strict mypy passed. Browser coverage includes save/load rollback, optimizer Apply, product undo, keyboard sizing/save, French locale, 320/390/640 px layouts, 200% CSS scaling, and default/assessed/multi-intake PDFs. The default Letter report is five pages instead of six. The affected print workflows were rerun after the final pagination fix.

Browser validation used Chromium; other engines and a human screen-reader session remain unverified.