fix(acp): support root session replacement in V2

[nullStack65]

What this is

Fork-held slice for the V2 architecture (USE_GENERIC_ACP_V2_WITH_SMALL_GENERIC_FIXES). It adds the generic, provider-neutral root-session-replacement capability the V2 ACP runtime needs so an agent that swaps its root session behind the same connection (omp /fresh) does not go silent.

• Upstream PR source: pingdotgg/t3code#2829
• Exact feat(orchestrator): introduce new orchestrator pingdotgg/t3code#2829 head used: 3d45b305632c72cbb8fd8a6dd5077b94c25d7eb5
• Base snapshot branch: orchestrator-v2-base-3d45b3056 (3d45b3056)
• Feature branch/head: feat/acp-root-session-adoption-3d45b3056 (9ba3a54d83c0fc4f8317ae605bd6a6a528213396)
• T3 main at completion: 4749035bda13b4b6260499caedbc0d69a2f60e6f
• Frozen bespoke reference (unmodified): pingdotgg/t3code#11973 @ 4f9419b72bb92bebc6669dd95f345e1ccdf4bced
• Not intended for upstream submission until the V2 architecture stabilizes.

Problem

An agent can replace the root session on the same ACP connection. Real omp 18.2.4 behavior, proven live:

• durable ACP session = A (returned by session/new, addressable, replayable via session/load);
• the user runs /fresh;
• every later session/update notification arrives under a new id B, including the direct response to the /fresh prompt;
• B is not independently addressable: prompts/cancel must keep using A, and session/list only ever shows A;
• V2 dropped B as a foreign session, so the thread went silent after /fresh.

Design (generic, no provider branches)

AcpSessionRuntimeOptions.adoptRootSessionReplacement (default false) plus onRootSessionReplaced for diagnostics.

• Durable vs live identity: the durable setup id is never rewritten. Prompts, cancellation, session/load, session/close, model/config requests, and assistant item identity all keep using the durable id.
• Adoption rule: a new id can replace the live root only when it is first seen while a root prompt is in flight, and at most once per prompt. Foreign ids seen while idle are always rejected (child/background traffic cannot become root accidentally).
• Projection: adopted notifications are projected back onto the durable id before either the runtime's internal consumer or the adapter sees them, so downstream session-id checks cannot reject the live root.
• Durable stragglers accepted, stale replaced-live ids dropped: updates on the durable id remain root traffic; a previously adopted live id is dropped once replaced.
• Wiring: the generic ACP Registry adapter exposes a per-instance opt-in through AcpRegistrySettings.rootSessionReplacement (default false, generic, usable by future agents). No global relaxation for every registry agent. No if agent === "oh-my-pi" anywhere.
• The opt-in contract is explicit: enable only for agents whose connection publishes a single live root session id and never child/subagent session ids; child traffic must be normalized before it reaches the runtime.

Net change: 6 files, ~+609/-5.

Tests

apps/server/src/provider/acp/AcpJsonRpcConnection.test.ts — new root-session-replacement block (all green, 47/47 total):

• A strict default: replaced root rejected (text stays root before fresh)
• B opt-in: A→B adopted, B updates continue, item ids stay in the durable namespace
• C/E successive replacements A→B→C deterministic; onRootSessionReplaced reports each change once
• D prompt/cancel requests keep the durable id after replacement
• F idle cancel while a replaced-root turn hangs targets the durable id (and never B)
• G durable stragglers accepted, stale replaced-live ids dropped
• H foreign session while idle never adopted
• I opt-in off = zero behavior change

Additional green:

• AcpRegistryAdapterV2.test.ts 3/3 — includes a new end-to-end test through the real registry runtime with the instance opt-in.
• AcpAdapterV2.test.ts 105/105 (V2 provider adapter runtime policy).
• OrchestratorReplayFixtures.integration.test.ts 73/73 (V2 replay fixtures).
• apps/server tsc --noEmit: clean. packages/contracts tsc --noEmit: clean.
• Targeted vp lint on every changed file: clean. git diff --check: clean.
• Test-fixture-only platform fix: the registry fixture did not declare a darwin-x86_64 distribution, so the pre-existing E2E test failed on this host. Same failure reproduces on the base snapshot 3d45b3056 (ACP Registry agent fixture-agent has no compatible distribution for darwin-x64), classified as a base/host fixture gap, now fixed in the fixture (no production code).

Live zero-inference proof (real omp 18.2.4, registry PR pingdotgg#613 metadata)

OMP updated to 18.2.4; local binary sha256 780a47a5... matches the exact darwin-x86_64 sha256 in agentclientprotocol/registry#613 (eed36bcaa677749b0bdf24010dfc3883834612c9). Registry JSON served from an injected HttpClient using the exact oh-my-pi/agent.json metadata; commandPath pointed at the local omp; authMethodId: "agent". No model prompt; only local commands (/context, /fresh, /rename).

Through the patched V2 registry adapter (rootSessionReplacement: true):

• durable native id after session/new: 01a0b151-096c-...
• /context fix(acp): support root session replacement in V2 #1 → Context window: 272000 tokens (6% used) (updates flow)
• /fresh → Fresh provider session started (1 provider state pruned).
• /context fix(server): opt-in fail-closed policy for codex thread resume #2 (post-fresh) → Context window: ... — updates continue through the patched runtime
• command catalog published; durable native id after /fresh unchanged: 01a0b151-096c-...

Direct AcpSessionRuntime with adoptRootSessionReplacement: true:

• start id A = 01a0b151-102d-...; session/list contains A
• /fresh → onRootSessionReplaced { previousSessionId: A, sessionId: 01a0b151-12be-... } (live identity observed without overwriting A)
• idle session/cancel succeeds on A
• /context after /fresh succeeds
• session/load(A) returns A; session/list after load still contains A

Dependencies / not covered here

• Registry metadata remains agentclientprotocol/registry#613 (oh-my-pi 18.2.4). This branch does not vendor registry metadata.
• Inference-gated items (model prompt round-trips, thinking/model config application, usage) were intentionally not exercised; this proof stops at zero-inference behavior.
• Separate follow-ups (not in this slice): ANSI cleanup, thread-title projection, subagent presentation, usage UI, onboarding.