Validate semantic preservation and assistive-technology behavior on supported documents

## Priority

P2 — focused reliability or product improvement

## Context

Portfolio audit follow-up (2026-09-08). Contrast/typography checks alone cannot establish comprehension or universal accessibility. Current format support is limited, so a useful next milestone is reviewable transformation of supported HTML/controlled templates.

## Implementation scope

Build a pinned document corpus containing headings, reading order, tables, lists, forms, links, numbers and deadlines. Add a before/after change report and a reversible original/output workflow. Use model suggestions only as candidate transformations with deterministic preservation checks. Add a manual assistive-technology protocol and clearly state unsupported PDF/OCR/CSS cases.

## Acceptance criteria

- [ ] Text, numeric values, deadlines, link targets and meaningful ordering have preservation fixtures.
- [ ] Unsupported structures are flagged for review rather than silently certified.
- [ ] The change report identifies what changed and retains access to the original.
- [ ] Screen-reader/keyboard checks are documented separately from automated contrast metrics.
- [ ] A consented reader-test protocol is available; measured results are not invented.
- [ ] Public claims match supported formats and do not promise guaranteed compliance.

Keep the change focused on this issue. Use synthetic fixtures for sensitive data and mocked providers for automated tests; do not require paid model calls to run the regression suite.

Start by locating the current supported HTML or controlled-template transformation flow and the existing contrast, reading-order, and alt-text checks. Build synthetic corpus fixtures and mocked-provider regression coverage, then verify preservation reports, reversible outputs, unsupported-case flags, and separately documented assistive-technology and consented reader-test protocols.

Written by the indexing model from the issue text.