feat(unstable-v2): add session injection schema - #2043

Sharing some live, empirical data that might be useful here, since this PR is where these semantics actually land in the schema.

We ran a controlled comparison of four CLI coding agents (Codex CLI, Claude Code, OpenCode, Google's Antigravity CLI) against the same steering scenario — a multi-step task, corrected mid-run with "stop after step N, don't do the rest" — and the results map fairly cleanly onto SessionInjectSteerInStream:

• Codex CLI (its own turn/steer RPC): behaves like interrupt — lands near-instantly, mid-generation.
• Claude Code: no dedicated steer RPC at all — writing a new user message onto the persistent --input-format stream-json stdin while a turn is running gets folded in live, but only at the next safe boundary, closer to finish than interrupt. This also confirms a concern @SteffenDE raised on the RFD thread (docs(rfd): mid-turn input via session/inject (queue and steer) #1261) about the Claude Agent SDK only generating a message ID at delivery time: there's no separate pending-ack step in the CLI's own mechanism at all, it's a raw stdin write — an ACP adapter really would have to synthesize InjectSessionResponse.messageId itself rather than getting one back synchronously.
• OpenCode: has genuine mid-turn steering through its own session API too, also finish-like — worth noting since neither OpenCode nor Antigravity come up anywhere in the RFD discussion.
• Antigravity CLI: no live steering mechanism at all, confirmed directly against the CLI, and no roadmap signal either.

For the two without live steering, our fallback measurements might be useful for gauging how costly the plain queue-after-cancel path actually is in practice: giving the resumed turn the actual correction as its real next prompt (not a generic "continue") correctly handles genuine redirection, not just abort — verified live. It's not free though: it pays for the full prior context again as fresh input, plus real wall-clock restart time (process teardown + reinit) — structural costs, independent of whether the caller times the cancel to land on a clean step boundary.

Happy to share more detail on methodology if it's useful for pinning interrupt/finish against real agent behavior.