OMP-4 — Prove ACP reliability parity for restart, cancel, process death, mobile, and runtime ownership #56

Parent epic: #21
Depends on: #54, #55
Parity contract: #52

Objective

Prove the ACP path is at least as reliable as the terminal/transcript path before making it default.

Turn ownership / cancellation

Test:

• Stop during active turn;
• repeated Stop;
• stale Stop after turn ends;
• cancellation while a tool or interaction is active;
• provider self/background activity;
• new user send after cancellation;
• generation/session replacement fencing.

A stale action must never affect newer work.

Process lifecycle

Through the merged managed-provider lifecycle:

• graceful provider exit;
• crash;
• forced kill;
• descendant/unverifiable exit cases;
• stderr drain/bounds;
• restart only after proven close;
• process dies during prompt;
• process dies during interaction.

Restart / replay

Simulate Orca restart with:

• settled turn;
• active/incomplete turn;
• incomplete tool;
• pending interaction;
• queued/admitted user send.

Assert the shared journal/replay lifecycle produces one coherent conversation and never silently resends ambiguous input.

Runtime/locality

Compare against #52's existing support floor:

• desktop local;
• headless/runtime-owned sessions;
• paired/mobile clients;
• Windows/macOS/Linux;
• WSL/SSH only where currently guaranteed.

ACP may expand locality. It must not silently remove a supported case.

Mobile

Verify OMP structured ACP sessions expose the same or stronger:

• history;
• send;
• options;
• interactions;
• status.

Acceptance criteria

• Reliability parity rows in OMP-0 — Freeze the terminal/transcript Native Chat parity contract #52 are green.
• No automatic resend after ambiguous delivery.
• Restart/replay is duplicate-free.
• Cancel/Stop is generation-safe.
• Process replacement waits for generic proven-exit semantics.
• Supported mobile/runtime cases are no weaker than before.