Bug: Fragment Divergence & Text Search Breakage After comment.add

[lpcode808]

When adding a comment via /api/agent/<slug>/ops with comment.add, the internal document state becomes corrupted with HTML annotations. Subsequent API operations (edit/v2, suggestion.add, find_replace_in_block) all fail with FRAGMENT_DIVERGENCE or ANCHOR_NOT_FOUND errors because text search can't find content that exists in clean markdown.

Steps to Reproduce

1. Create a Proof document with text: "Hello whwat is this for"
2. Add a comment anchoring to that text:

curl -X POST "https://www.proofeditor.ai/api/agent/<slug>/ops" \
  -H "Authorization: Bearer <token>" \
  -H "X-Agent-Id: claude-code-haiku" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "comment.add",
    "by": "ai:claude-code-haiku",
    "quote": "Hello whwat is this for",
    "text": "This is a comment"
  }'

3. Attempt subsequent edit with /edit/v2:

curl -X POST "https://www.proofeditor.ai/api/agent/<slug>/edit/v2" \
  -H "Authorization: Bearer <token>" \
  -H "X-Agent-Id: claude-code-haiku" \
  -d '{
    "by": "ai:claude-code-haiku",
    "baseRevision": 2,
    "operations": [
      {"op": "insert_after", "ref": "b1", "blocks": [{"markdown": "New content"}]}
    ]
  }'

Result

FRAGMENT_DIVERGENCE error. Snapshot shows HTML annotations in markdown: <span data-proof="comment" ...>Hellowhwat is this for</span>

Root Cause

Comment annotations corrupt the canonical markdown representation. Text search fails because it searches plain text but snapshot returns annotated HTML.

Impact

• Cannot chain comment + edit operations
• Document becomes read-only for programmatic edits after comment added
• Breaks agent workflows mixing comments and content changes

Environment

• Hosted Proof (proofeditor.ai) - 2026-03-11
• Verified with /api/agent/ endpoints

[dshipper]

Investigated this in every-proof on March 11, 2026.

Current status:

• I could not reproduce the exact hosted corruption path locally on current server code. In local repros, comment.add followed by /edit/v2 and suggestion.add succeeds, and both /api/agent/:slug/state and /api/agent/:slug/snapshot stay span-free.
• I still bundled a related hardening fix because the reported symptom is adjacent to a real gap: agent-facing /state, /snapshot, and share JSON/markdown now strip all Proof spans, not just authored spans, so comment/suggestion annotations cannot leak into agent-readable markdown.
• I also found and fixed a separate persisted-Yjs replay bug while validating this: on a fresh DB, the first update after a snapshot could be skipped during hydration, which let a later marks-only write roll markdown back to stale content. That was not the same bug as the issue report, but it was real and showed up in the same validation pass.

Added regression coverage for:

• browser-authored comments being readable immediately through state.marks
• comment.add not contaminating agent-readable markdown or breaking find_replace_in_block / suggestion.add
• on-demand repair docs staying writable after edit + comment

Verification on the branch:

• npm run test:server-routes-share passes

So my read is: this issue is related, but the exact hosted failure likely involves either an older deployed build or a hosted-only path I couldn’t reproduce locally. The current branch still meaningfully hardens the affected surfaces and closes a nearby persistence bug.

[devYRPauli]

Two related fixes/areas for this:

• The "internal document state corrupted with HTML annotations" part: that is proof/authored <span data-proof=...> markup ending up on the canonical row, which the text-search/verification path then cannot reconcile against clean markdown. PR fix(agent-edit): stop replace from orphaning a closing </span> on wrapped authored spans #59 fixes one case where a replace over a wrapped authored span orphans the closing </span> (expandRangeToIncludeFullyWrappedAuthoredSpan in server/proof-span-strip.ts only expanded on exact content-bound equality), leaving malformed span HTML on the row. It is probably not your exact comment.add path, but it is the same root area: proof spans in the canonical markdown that downstream ops then trip over.

• The subsequent FRAGMENT_DIVERGENCE on edit/v2: PR fix(agent-edit-v2): detect live ref drift against persisted base not live fallback #62 addresses one cause of that, where the live ref-drift guard never fires under live presence (it compares the live fragment against itself). Different trigger from yours (topology drift vs annotation corruption), but the same failure surface.

Happy to dig into the comment.add -> annotation path specifically if that would help.