fix(composer): use native ProseMirror inserts for autocomplete (mentions / channels / emoji)

[tlongwell-block]

What

Replace the markdown round-trip on mention/channel/emoji autocomplete with a single native ProseMirror transaction, and fix a latent hardBreak off-by-one in the plain-text cursor mapping while we're there.

Tracked the design with Perci end-to-end in #sprout-bugs; we converged on 9.5/10 (minimal / elegant / correct) before I started building.

The four bugs and what's actually wrong

Reports in #sprout-bugs:

1. List toggle wraps the whole message (Thomas P) — toolbar UX gap. Not in this PR — see "Follow-up PR" below.
2. Newlines collapse when you add an emoji or @-tag. Root cause: the autocomplete path calls richText.setContentWithTrailingSpace(nextContent) → editor.commands.setContent(markdown). That's a full markdown reparse on every insert; even with breaks: true, the round-trip is sensitive to subtle parser/serializer asymmetries and resets cursor/marks.
3. @mentioning ruins the formatting. Same root cause — the reparse rebuilds the doc from scratch, so any in-progress marks (bold, italic, code) and list-item nesting at the cursor are blown away.
4. Adding links is wonky (Thomas P) — toolbar UX gap. Not in this PR — see "Follow-up PR" below.

This PR fixes #2 and #3 plus a latent off-by-one in getTextAndCursor: it used state.doc.textContent, which ignores Tiptap's hardBreak renderText. Every Shift+Enter shifted the cursor offset by 1 relative to what mention/channel detection saw — invisible only because nobody had to round-trip the offset back to a PM position.

How

1. New module plainTextProjection.ts — one walk over the doc, one source of truth for both directions:

buildPlainTextProjection(doc) → {
  text: string,
  mapPMToTextOffset(pm)  → number,
  mapTextOffsetToPM(off) → number,
}

Hard breaks and inter-block boundaries each contribute exactly one \n, matching doc.textBetween(0, end, "\n", "\n"). Forward + reverse mappings share the same segment list so they can't drift — historically the source of bugs.

2. useRichTextEditor gains two paired helpers:

getPlainTextAndCursor()      // replaces getTextAndCursor
replacePlainTextRange(from, to, text)

replacePlainTextRange is a thin wrapper around tr.insertText(text, fromPM, toPM) — the right ProseMirror primitive for a textarea-style replacement (preserves active marks, handles empty insert). The new cursor lands at tr.mapping.map(fromPM + text.length) so the helper stays robust if anything else shifts positions.

3. Hook API change:

// before
insertMention(suggestion, content, selectionEnd)
  → { nextContent: string, nextCursor: number }

// after
insertMention(suggestion, selectionEnd) → AutocompleteEdit
//   { replaceFromOffset, replaceToOffset, insertText }

Hooks stay editor-agnostic: they describe the edit in plain-text-offset space. The editor wrapper applies it. Same change for useChannelLinks and useEmojiAutocomplete.

4. Both composers (MessageComposer + ForumComposer) swap to the new helper via a small shared applyAutocompleteEdit local. No more setContent(markdown) for cursor edits.

Removed

• setContentWithTrailingSpace (no callers).
• The trailing-space ProseMirror hack (insertText already ends in a space; no markdown reparse to evade).
• The old descendants walk in getTextAndCursor (replaced by the projection helper, which is correct for hardBreak).

Preserved

• breaks: true on tiptap-markdown (already set).
• \\\n → \n strip in getMarkdownFromEditor — wire-format behaviour, no longer load-bearing, safe to leave.
• Emoji-picker button + toolbar @ button paths — already native chain().insertContent(); control cases, unchanged.

Tests

• New plainTextProjection.test.mjs — 20 tests against the real Tiptap schema (built via getSchema([StarterKit.configure(...)]), not a hand-rolled one): single paragraph, hardBreak before/after/across-break, two and three paragraphs, paragraph + bullet-list mix, list-item round-trip, edge cases (empty doc, out-of-range), and a property test that every text offset → PM → text offset is identity over a fixture covering all three node types.
• Removed useRichTextEditor.test.mjs — it tested a copy of the old algorithm against a hand-built node list. New test exercises the real schema directly.
• 176/176 existing node tests still pass; typecheck + biome + production build + full pre-push CI suite all clean.

File-size and LOC

• MessageComposer.tsx: 702 → 709 lines (limit 710; fits).
• Existing files: +155 / -273 (net −118).
• New files: ~200 LOC pure projection + 220 LOC tests.

Self-rating

• Minimal — no dependency changes; one new module + one named type; deletes more than it adds.
• Elegant — one plain-text projection model; hooks describe edits, editor applies them; ProseMirror knowledge lives in one file.
• Correct — fixes both reported bugs and the latent hardBreak offset bug; eliminates the reparse class of problems regardless of which exact parser asymmetry is currently biting.

Follow-up PR (separate, after this lands)

• List toolbar (Dependency Dashboard #1): collapsed cursor in non-empty paragraph → split into a new empty block then toggleBulletList.
• Link toolbar (docs(readme): clarify desktop setup #4): replace window.prompt × 2 with a shadcn popover; setLink({href}) for selections, native insertContent with link mark for no-selection — never insert literal […](…) text.

These are toolbar UX changes on a different surface; reviewing them mixed in with this deeper editor-state change would slow review.

Out of scope

• Draft attachments lost on channel switch (useDrafts.ts) — separate file/feature.
• Library migration (Lexical / CodeMirror) — we ruled this out in the design thread. The bugs are local code patterns; we'd reproduce them in any editor by following the same string-mirror approach.

---

cc @perci-sprout — your two implementation nits from the design thread are both honored:

1. Cursor PM after insert goes through tr.mapping.map(fromPM + insertText.length), not the raw arithmetic.
2. Tests explicitly cover before-/after-/across-<br> and inside-list-item.