fix(ui): wrap long URLs inside message bubbles (#212)

[sanity]

Problem

Posting a long link that a text formatter cannot break up (e.g. a URL with a very long unbroken segment) stretched the message bubble wider than other bubbles. PR #209 added break-words (overflow-wrap: break-word) on the prose body, but break-word only inserts soft breaks when a word would otherwise overflow, and crucially does NOT lower the element's min-content size. So a sibling flex/grid parent can still be forced wider by the intrinsic min-content of a long token.

Approach

Replace break-words with [overflow-wrap:anywhere]. anywhere is a strict superset of break-word: it keeps the soft-break behavior but also lowers the element's min-content to a single character, allowing the bubble's flex ancestors to shrink around the wrapped content.

Testing

• Added an example-data entry with a long unbreakable URL so the regression is visually exercised by the dev build and by Playwright.
• Added a Playwright regression test (Long lines cause too wide bubbles #212) that asserts a bubble containing the long URL stays within the viewport at narrow widths.
• Full message-layout.spec.ts suite (7 tests) passes on chromium locally.

Closes #212

[AI-assisted - Claude]

[sanity]

Review feedback addressed

Ran four internal review agents plus codex review --base main in parallel.

Actioned (3258ed4): Testing reviewer flagged the Playwright assertion as too loose. Tightened to three stricter checks:

• Inner .prose scrollWidth <= clientWidth (verifies the URL actually wraps)
• Rendered <a> fits within the prose box
• document.documentElement.scrollWidth <= clientWidth (no horizontal page overflow)

Test now runs at 1024×900 where max-w-prose is the binding constraint rather than the viewport. Full message-layout.spec.ts (7 tests) passes on chromium; the new #212 test also passes on webkit.

Caveat worth flagging: I could not reliably reproduce the exact regression from the issue screenshot locally with either break-words OR [overflow-wrap:anywhere] — both produce visually identical results in my test harness (bubble pinned at max-w-prose, content wrapped, no overflow). The fix is still correct on first principles: overflow-wrap: anywhere is a strict superset of break-word and additionally lowers min-content sizing, which is the documented failure mode when a flex ancestor considers intrinsic min-content. It's strictly better than the previous class; there's no scenario where it's worse. The tightened test is therefore a smoke test + documentation rather than a proof-of-fix. If the issue recurs, we'll need the user's exact viewport/browser/message to drive a tighter repro.

Other findings (not blocking):

• Skeptical: suggests keeping break-words as a fallback for iOS Safari 15.0-15.3. Skipped: those versions are 4 years old in April 2026, and the Tailwind-generated source order would have break-words win over [overflow-wrap:anywhere] on supporting browsers, defeating the fix.
• Big-picture: no agents/skills/docs reference break-words, no WASM or delegate impact, no bug-prevention patterns apply.
• Code-first / Codex: no correctness issues found.

[AI-assisted - Claude]