Add ui-test skill: bugs before beauty before drop-dead gorgeous (CL-6650)

[TheGreatAxios]

Summary

• Adds .claude/skills/ui-test/ — a repo skill any lane can invoke to grade a live UI surface through the owner's progression bugs → beauty → drop-dead gorgeous, per CL-6650.
• Studied browserbase/skills' ui-test and kept: the adversarial stance, before/after snapshot evidence discipline, structured pass/fail reporting, and the reference-file-on-demand structure. Changed: browse → agent-browser as the driver (named sessions, @eN refs, scoped snapshots), generic UX heuristics → our three CL-6650 tiers (each blocking the next), and "find any bug" → "stop at the first tier failed."
• Tier 1 (bugs, blocks everything): entity sanity, names in name slots, no placeholders, no dead buttons, ≤100ms feedback, failures name a refId.
• Tier 2 (beauty, blocks demo-path surfaces): identity per row, DESIGN.md spacing, consistent loading treatment, consumer language.
• Tier 3 (gorgeous): the Vercel/Stripe/Apple side-by-side judgment, argued with evidence rather than asserted.
• Calls out the stale-match trap explicitly: never grep a whole-page snapshot (sidebar bench previews echo strings that also appear in the timeline/dialog under test) — always scope with snapshot -i -s <selector>.
• .gitignore: added a !.claude/skills/ exception (via .claude/* + re-include) so this repo skill is tracked — .claude/ was previously blanket-ignored as local orchestration noise, which conflicted with CL-6650 wanting a committed, lane-invokable skill.

Acceptance run (this PR)

Ran the skill by hand against the New Workbench prompt-first screen on :3000, logged in as quinn.baker@example.com, session uitest.

## UI Test — New Workbench prompt-first screen

**Tier reached:** 3 (gorgeous, with notes)
**URL:** http://localhost:3000 → New workbench
**Session:** uitest

### Blockers
(none)

### Tier 3 notes
- What already clears the bar: tight spacing, icon + name + one-line
  consumer description per template card ("Code review", "Due Diligence",
  "Just start talking"), an orange-ring-focused prompt input that mirrors
  DESIGN.md's accent, and the "Takes about ten seconds" line — consumer
  language, not system internals. Typing into the prompt box enables
  "Start this workbench" with no perceptible delay (scoped snapshot
  before/after, ref e22: disabled → enabled on the same snapshot call
  that followed `fill`). Clicking "Just start talking" (ref e17)
  navigated straight into a live bench with Myra's greeting already
  rendered — no "is it working?" gap.
- One step short: the sidebar lists two sibling benches both titled
  "New Workbench," distinguished only by preview snippet and timestamp
  (scoped snapshot `-s ".shell-main"` was clean; this was visible on the
  unscoped sidebar list, ref e17 vs e19 pre-navigation). Not a Tier 1/2
  blocker — it's default naming, not wrong data — but a gorgeous-tier
  surface would rename a bench from its opening prompt/template choice
  sooner than "New Workbench" persisting after the first message lands.

### Passed checks
- Tier 1 — entity sanity — sidebar list contains only workbench rows, no
  routines/agents mixed in (full snapshot, refs e17–e19).
- Tier 1 — names in name slots — template card titles ("Code review",
  "Due Diligence", "Just start talking") never render their description
  text (scoped snapshot `-s ".shell-main"`).
- Tier 1 — no placeholder content — no seeded/lorem copy anywhere on the
  screen.
- Tier 1 — no dead buttons — all three template cards and the prompt
  submit path are wired; "Just start talking" (e17) navigated to a real
  bench.
- Tier 1 — feedback ≤100ms — "Start this workbench" (e22) flips
  disabled→enabled in the snapshot immediately following `fill`.
- Tier 2 — spacing/identity — icon + name + description per card, tight
  spacing, no blank-pane loading gap observed.
- Tier 2 — consumer language — "Tell it what you're trying to get done.
  Takes about ten seconds," "An empty room with Myra in it" — no system
  internals in copy.

CL-6649 seeded-regression follow-up

CL-6649 (invite-agent dialog leaking routines in as agents) has not landed yet (no matching commit or PR on main as of this branch). Per the task's guidance, not blocking on it — the seeded-regression acceptance check (temporarily rename Jimmy's row to his description, confirm the skill flags it) is follow-up work once CL-6649 merges. references/regression-seeding.md documents exactly how to run that check when it's ready.

Test plan

• bun run lint — 0 errors (10 pre-existing warnings unrelated to this change)
• Ran the skill live against :3000's New Workbench prompt-first screen as quinn.baker@example.com via agent-browser --session uitest; report above
• Follow-up once CL-6649 lands: seed the regression from references/regression-seeding.md and confirm the skill flags it