fleet(#1376): real-ssh two-engine e2e test — AC2/AC3 proof, CI-gated #1383

Acceptance Criteria

• Two engine instances run on distinct local ports during the test
• An ssh -L forward connects the test client to the "peer" engine
• Engine + SSE requests through the resolver land on the attached peer's engine (verified by a distinct response marker)
• /amicode/roster requests land on the keeper (verified by a distinct response marker)
• The test is CI-gated (runs in the main test suite, not opt-in)
• p95 attach-to-first-response ≤ 2000 ms (AC4 bar from Peer fleet studios + compute federation (ADR 0027) #1340)

Testing Decisions

New integration test file (this IS the test). Model on the AC4 p95 harness from #1353 which already ran over real ssh -N in CI. The two engines can be lightweight mock HTTP servers that return distinct markers — they don't need to be full opencode instances.

Key Decisions

• Real SSH, not mocked. AC3 is explicit: "not a stub." ssh -L to localhost is the minimal real-transport path.
• CI-gated, not nightly. The Peer fleet studios (ADR 0027, Horizon 1) — #1341–#1346 #1353 p95 harness proved ssh-to-localhost is stable in CI (~165 ms).

Constraints & Invariants

• The test must clean up SSH processes on teardown (no orphaned ssh -N children).
• The test must not depend on fleet enrollment state — it sets up its own lightweight fixture.

Prior Art

• The Peer fleet studios (ADR 0027, Horizon 1) — #1341–#1346 #1353 AC4 p95 harness (amicode_service_attach_switch_latency.test.ts).
• The attachment_transport.ts ssh -L forward.

Source

Part of #1376. Blocked by #1378 (FleetPlane + resolver wiring) and #1381 (attach/detach lifecycle).