Flip workbench creation: prompt box first, prefabs underneath

[TheGreatAxios]

Summary

Implements CL-6628: flips the New Workbench screen's hierarchy from
"choose a kind, then Create" to "say what you want, or pick a shortcut."

• Prompt box on top, autofocused, placeholder "What do you want your
  Workbench to do?" Enter (or the send button) creates a blank
  workbench and delivers the typed text as the person's own first
  message via sendMessage/partsForSend from @corbits/chat-ui, so
  Myra's next line responds to the actual intent, not a kind label.
  Delivery happens after the template/setup greeting posts and before
  navigation, so the goal lands as the next message in the room
  rather than racing the greeting.
• Prefab cards (Code review, Due Diligence, Just start talking) below
  the prompt box — one click creates immediately through the same
  createWorkbenchFromTemplate / instantiateWorkbenchTemplate paths
  as before. No radio-then-Create two-step anywhere on the screen.
• Fixes the ~300ms blank-pane flash on create (CL-6623 finding Bench chat defaults with per-channel context-window overrides #3):
  the creating branch now renders WorkbenchLoadingState with
  delayMs={0} instead of the default 200ms hold-back, since the
  moment Enter/a card is clicked is already a known-genuine wait, not
  a maybe-fast one.
• Dropped the disabled "More kinds soon" row along with the
  radiogroup — three prefab cards, no dead entry.

react-ui-drift's snapshot count moved 48 → 49 for the prompt box's
<textarea> — react-ui has no Textarea primitive, and this matches
every other multi-line input in the codebase (composer.tsx,
form-block.tsx, general-section.tsx, …), so it's the same convention
carried forward rather than new drift needing a fix.

Live verification

Ran the app against a real hub+sidecar stack (own .env, isolated DB
so as not to touch another lane's running dev stack) signed in as
dana.reyes@example.com, with an Ollama credential planted against
the owner's remote tailscale box so Myra could actually reply — no
local inference was run.

• Typed "Get our onboarding docs into shape for new hires" into the
  prompt box, hit Enter: loading state showed immediately (no blank
  frame), the room opened with Myra's setup greeting followed by the
  typed goal as dana's own message, and Myra started a real reply
  turn. (The reply itself 401'd against Anthropic in that ad hoc
  verification environment — a model-routing/credential concern in my
  throwaway stack, unrelated to this change; the message delivery and
  turn kickoff both worked.)
• Clicked the Due Diligence card: same immediate loading treatment,
  no blank flash, and the room opened with Myra + Scout in the roster
  and the template's own greeting — a single click, no intermediate
  step.

Test plan

• WORKBENCH_CHECK_SINCE=origin/main bun run typecheck — clean
• WORKBENCH_CHECK_SINCE=origin/main bun run lint — clean
• WORKBENCH_CHECK_SINCE=origin/main bun run test — 849 pass, 0 fail
• bun run check:structural — all checks pass (including
  react-ui-drift after the snapshot bump)
• Rewrote apps/web/test/new-workbench-picker.test.tsx and
  apps/web/test/toast-single-system.test.tsx for the new prompt
  box + prefab-card shape
• Manual browser verification (see above)

Note on scope

This branch only touches apps/web (plus one snapshot-count line in
scripts/checks). No changes to packages/chat-ui — the other
in-flight lanes there (failed-turn-strip, retry wiring, timeline perf)
are untouched.