feat: Gemini Enterprise Demo Generator v11.91-public

[ryotat7]

Description

Ports the Gemini Enterprise Demo Generator sample forward from v11.62-public
to v11.91-public. One new file (test_press_retire.py), no layout change —
the same sample, updated.

The largest change is a client scroll behaviour that made every A2UI button
press throw the reader backwards, and what the agent can do about it.

A2UI press landing

On a press, Gemini Enterprise scrolls to the element of the user's PREVIOUS
press. The test that settled it changed nothing in the agent: press a button
in the newest turn, then scroll up and press one in a much older turn, and the
view jumps down, onto the surface pressed a moment earlier. Direction is the
tell — no rule about "the nearest surface above the pressed one" can scroll
downward, so the earlier readings are withdrawn.

• That target is client state, and nothing the agent emits during the turn can
  move it. What it can do is delete what the target points at:
  _pressed_surface_delete_parts() reads userAction.surfaceId off the press
  payload and appends a deleteSurface as the last part of the artifact, so the
  anchor the next press would aim at no longer exists. Live-verified on a
  deployed demo: the backwards jump stops.
• The choice stays visible — a press also carries query.text, which the client
  renders as the user's own message. A surface the same turn is rendering is
  never deleted, and A2UI_KEEP_PRESSED_SURFACE=1 restores the old behaviour.
• test_press_retire.py exercises the helper off-line (chip, gate and welcome
  presses, the re-render guard, typed messages, malformed payloads, the kill
  switch), because the helper deletes a surface the user can see.
• Buttons still go in a trailing, card-rooted surface of their own — not as a
  scroll fix any more, but because next actions below the answer card is the
  layout the demos want. The server-built gate cards keep their prose in the
  card and move their input fields out together with the buttons: a
  {"path": ...} binding only resolves inside the surface the button lives in,
  so the fields cannot stay behind. Compose, confirmation and what-if cards keep
  their buttons inside, and so does the welcome card.
• Eight few-shot examples move their follow-up row out to the trailing surface,
  and the prompt no longer asks every card for a footer action row.
• Component properties the v0.9 composite catalog does not declare
  (MaterialRow.ariaLabel and friends) are removed: the renderer is strict where
  draft-07 validation is not, and one foreign property drops the whole surface.
• AGENTS.md section 2.6 records the rule, the test that pinned it, and the
  dead ends (an anchor surface before the card, an invisible trailing landing).

Registration survives a broken authorization

The setup script warned when the authorization write failed and then handed
$AUTH_ID to register_agent.py anyway. That is worse than it sounds:
Discovery Engine resolves authorizationConfig.agentAuthorization at create
time and rejects the whole registration with 404 NOT_FOUND when it names
nothing — so a broken authorization did not cost the demo its Workspace tools,
it cost the demo its agent and its direct chat link.

• Read the authorization back. A 200 on the POST or PATCH is not proof the
  resource is readable, and only the read answers the question the registration
  actually asks. Not readable → clear $AUTH_ID, register without it, and say
  plainly what was lost.
• Retry the registration once without it. Reading back fine is still no
  guarantee the registration accepts it (bound to another agent, wrong project
  number). A registered agent with degraded Workspace tools beats no agent at
  all.
• Both call sites — the single-app branch and the pick-an-app branch — now go
  through one register_ge_agent_with_fallback, so the two copies cannot drift
  apart again.

Prompt cost

• The setup script prunes the seven Workspace few-shot examples when Workspace
  MCP is off. a2ui globs the examples directory, so they were prefilled on every
  turn — about 5.5k tokens teaching an agent to compose a Gmail draft it has no
  tool to send. A2UI_KEEP_ALL_EXAMPLES=1 keeps them.
• The BigQuery toolset is filtered to the three tools the instruction actually
  uses, and the routing block is rewritten so a figure question is one
  execute_sql call instead of a metadata expedition in front of it.

Hardening in the generated demo

• Auth-token fingerprints in the logs instead of token prefixes, in both the
  agent and the A2A runtime.
• sha256 for the idempotency key. Task failures and the viewer app's error
  responses return a generic message and log the detail — these run on public
  Cloud Run URLs and the exception text carries Firestore paths and project
  identifiers.
• Record and task ids are restricted to identifier characters before they reach
  an inline handler, the goal preview escapes its markdown input, and the two
  version-pinned CDN scripts carry subresource integrity hashes.

Sample data and defaults

• The few-shot business data is genericized: no real company names, English
  throughout, so the examples do not steer the agent's output language.
• Default agent model moves to Gemini 3.7 Flash.

Validation

• All 26 A2UI example JSONs parse and validate against the v0.9 composite
  catalog (validate_examples.py), and the 9 template Python files compile.
• python3 test_press_retire.py — 9 cases, all passing.
• node --check on Code.gs; bash -n on all six renders of the setup script
  (managed-agent × Workspace-MCP × Workspace-auth), with the authorization
  read-back present in exactly the four Workspace combos and the registration
  fallback used at both call sites in all six.
• Nine words added to .github/actions/spelling/allow.txt.

Checklist

• Follow the CONTRIBUTING Guide.
• Ensure the tests and linter pass.
• Appropriate docs were updated (AGENTS.md).
• The author is listed in CODEOWNERS for these files.