A stored image authored without `alt` gained a phantom `alt: ""` on every
round-trip (`markdownToProseMirror(convertProseMirrorToMarkdown(doc))`): `marked`
renders `` as `<img alt="">`, and the stock tiptap Image `alt` parseHTML
(`getAttribute("alt")`) materialized the empty string where the original had no
attribute. That false diff is a real GS-EDIT-REVERT churn source — an agent /
git-sync touch of a page with an image mutates the stored JSON (`absent -> ""`),
producing phantom diffs that can overwrite live edits.
Fix is PARSE-SIDE ("" ≡ absent), so the RAW round-trip is idempotent — not only
the canonical form (history / stored JSON diff on the raw shape; masking it only
in canonicalize would leave that noise). `image.alt`/`title` parseHTML now coerce
`getAttribute(...) || null`, plus defense-in-depth `|| null` across the at-risk
empty-string class (video aria-label, drawio/excalidraw title+alt, pdf name,
attachment name+mime) matching the existing `image.caption || null` precedent.
NOTE — image `align` is NOT changed: it round-trips correctly (center via the
schema default "center", left/right via the `<!--img {...}-->` comment). Its
`toBeUndefined()` in the git-sync gate is canonical-form normalization, not a loss.
Intentional divergence from editor-ext: editor-ext's literal `alt` parseHTML
returns "" verbatim, but this coercion CONVERGES on editor-ext's real STORED
shape (an image inserted without alt has no `alt` attribute -> re-parses absent,
never ""), so the round-trip is idempotent and matches real documents.
Adds a reusable, node-agnostic round-trip-stability matrix helper
(test/roundtrip-stability.helper.ts) — given a node + attr spec it enumerates
default/non-default combos and asserts byte-stability of BOTH the raw and the
canonical round-trip (the documented numeric width/height→string coercion encoded
as an explicit allowed normalization) — driven over image + the whole media
family (video/audio/pdf/attachment/embed/drawio/excalidraw). The only raw
empty-string instability it found was image.alt; the family was already stable.
Gate: package suite 33 files / 672 tests passed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
agent_coder
2ce672709a