Adopt upstream Antigravity provider (#9348) without regressing Prime Agent

[rynfar]

Upstream T3 Code added a Google Antigravity provider via the official ACP agent.
It is the last unadopted commit from the 2026-09-03 review batch, deliberately
held back from PR #261 because of what it touches rather than its size.

Upstream commits

• 06336460c9988f29c71e839c4c9c840c4552e077 — feat(providers): add Google Antigravity via the official ACP agent (#9348) (169 files, 21,470 insertions)
• 1e05187309... — fix(antigravity): refresh the model manifest so older Gemini models fold as legacy (#9397)
• 2120fbc185... — fix(web): avoid duplicate Antigravity install status (#9419)

61 files conflict against pylon as of the batch-2 branch.

Why this is isolated

The new provider files are the easy part. The risk is that #9348 rewrites the
shared ACP runtime Pylon's Prime Agent depends on:

File	Change
apps/server/src/provider/acp/AcpSessionRuntime.ts	+469 / -?
apps/server/src/provider/acp/AcpCoreRuntimeEvents.ts	modified
apps/server/src/provider/acp/AcpRuntimeModel.ts	modified
packages/effect-acp/src/client.ts, protocol.ts, _internal/shared.ts	modified

PrimeAgentAcpSupport.ts, PrimeAgentDaemonManager.ts and
PrimeAgentDaemonAdapter.ts all ride on that runtime, and upstream has no
Prime Agent, so no upstream test covers the interaction.

This is not hypothetical. Three upstream commits in this batch broke Prime Agent
in ways upstream CI could never catch:

• #8337 treated any default instance without a legacy providers entry as
  deleted. Prime Agent is exactly that, so it went dark — not enabled, not
  available, gone from the model picker.
• #9164 stopped dispatching descriptor defaults as overrides, which also cut
  Prime's thinking-level defaults out of background text generation.
• #9154 extracted sendInitialize for a new health probe and bypassed
  runStartupRpc, Pylon's startup RPC timeout — the probe could hang forever
  against a wedged agent.

All three were caught only by Pylon's own tests or by reading the diff.

Work

1. Cherry-pick #9348, then #9397, then #9419, oldest first.
2. Resolve AcpSessionRuntime.ts and packages/effect-acp Pylon-first.
   Specifically confirm Pylon's runStartupRpc timeout still wraps every
   startup RPC, including any new probe path.
3. Check the new AcpSessionRuntime shape against PrimeAgentAcpSupport,
   PrimeAgentDaemonManager, PrimeAgentDaemonAdapter and
   PrimeAgentRecoveryLedger.
4. Rename user-visible "T3 Code" strings in the adopted settings and docs copy
   (#9348 carries several, e.g. "Antigravity is disabled in T3 Code settings").
5. Decide whether Pylon ships Antigravity enabled by default, and whether it
   belongs in DEFAULT_SERVER_SETTINGS.providers.

Verification

• Full server suite, not focused tests. Prime Agent regressions in this batch
  surfaced only in the full run.
• apps/server/src/provider/acp/ and apps/server/src/provider/prime/ suites.
• A real-client check that a Prime Agent thread still starts, streams and stops,
  since the shared session runtime is being rewritten underneath it.

Context

• Batch-2 adoption PR: upstream: adopt the 2026-09-03 T3 batch, pass two (in progress) #261
• Batch-1 (merged): upstream: adopt the clean half of the 2026-09-03 T3 batch (65 commits) #260
• Review ledger: .agents/upstream-review.md