fix(desktop): stop mention chips from swallowing the caret

[mattbalza]

What

Clicking just after a mention in the composer put the caret inside the handle, so the next keystroke turned @bob into @bobx.

A mention chip is an inline ProseMirror decoration over live, editable text, so its painted box doubles as a caret hit target. --inline-chip-padding-inline (0.25rem, markdown.css) paints 4px of chip past the last glyph. A click in that band is visually after the name but resolves to the document position before the trailing space.

Why not the obvious fix

The composer's contenteditable carries MESSAGE_MARKDOWN_CLASS, so it inherits .message-markdown .mention-chip { display: inline-flex; ... }. The tempting fix is to unset that. Measured against a live composer chip, it changes nothing:

candidate	trailing band	click 0.5–1px past the glyph
baseline (inline-flex, padding: 3px 4px 2px)	4.0px	caret INSIDE the chip
display: inline; padding: 0 2px	2.0px	caret INSIDE the chip
display: inline; padding: 0 4px	4.0px	caret INSIDE the chip
padding-inline: 0 (either display)	0.0px	past the box — caret lands in the space

The trap band is exactly the inline padding. inline and inline-flex hit-test identically.

The change

One declaration, scoped to the editable surfaces:

.rich-text-composer .tiptap {
  --inline-chip-padding-inline: 0;
}

Zeroing the variable rather than the padding keeps the agent chip's derived geometry consistent — its padding-left stays icon + gap and the ::before robot icon simply moves flush to the chip's left edge.

Rendered messages are untouched: they are not editable, so the band costs nothing there and the padded pill is the intended look.

Test plan

New E2E in desktop/tests/e2e/mentions.spec.ts, verified red on main (band = 4, expected ≤ 0.5) and green here. It asserts both halves:

1. the structural invariant — the chip's box may not extend past its last glyph;
2. the behaviour — a click 2px past the glyph followed by a keystroke does not weld onto the handle.

The behavioural half alone is not sufficient: it sits one pixel from the genuinely ambiguous end-of-word click that every editor resolves into the word, so the invariant is what pins the regression.

• pnpm lint clean, tsc --noEmit clean
• mentions.spec.ts + team-mentions.spec.ts + messaging.spec.ts + channel-composer-overflow.spec.ts: 99 passed. The one failure (messaging.spec.ts → "shows your avatar when profile avatar is set") reproduces unchanged on main and is unrelated.