Op::SendMessage god-payload → TurnSpec, and prove the UI never awaits send() on the input path

[Hmbown]

From the 0.9.14 refactor backlog.

Op::SendMessage (ops.rs ~109+) carries ~20 fields incl. Box<ResolvedRuntimeRoute>, Box<CompactionConfig>, hook executor, provenance. Extract struct TurnSpec so new per-turn authority fields stop accreting into the enum arm; every ACP/app-server fix (#6088, app-server client issue) touches this anyway.

Also: ENGINE_OP_CHANNEL_CAPACITY = 32 (engine.rs ~90). If anything on the poll thread send().awaits into a full op channel, input wedges while the engine is mid-turn. handle.rs exposes async send — trace the call chain from the event loop. If it is already try_send + error toast, close this half with a comment.

Size: S + investigation.

[Hmbown]

Triage (v0.9.14 run, partially verified — not in #6161): the input-path half is already correct for SendMessage — dispatch.rs:844 reserves capacity off the render thread and hands off with send_reserved_op (:887), with the invariant written in the comment. Residual send().await sites to audit/convert: handlers.rs:2086 (CancelSubAgent), :2570 (SyncSession). The 20-field god-payload → TurnSpec is the large slice and should follow the audit's conclusion.

[Hmbown]

Half landed on v0914-chunk2 (PR #6175): the input path no longer awaits a full op channel — 4d70e050 keeps send() off the UI thread's poll path. The Op::SendMessage → TurnSpec extraction is still open: ~136 construction/match sites, mostly tests.

[Hmbown]

TurnSpec extraction landed on v0914-chunk2 (commit e9de8a7, will land with PR #6175): both Op::SendMessage variants (engine core::ops + serializable protocol::op) now carry Op::SendMessage(TurnSpec). The internally-tagged newtype keeps the wire shape byte-identical ({"kind":"send_message",...}) — protocol round-trip tests confirm. handle_send_message takes the spec, dropping the 23-arg signature. Also converted the audit's one missed input-path site: agent-focus FollowUpSubAgent in dispatch_composer_message now uses try_send + busy/unavailable reason instead of send().await. Verified: check -p codewhale-tui --all-targets clean; protocol 17/17; nextest core:: 837/837, runtime_threads 206/206, runtime_api 239/239, exec_agent 9/9, agent_focus 11/11 (2 tui::ui layout tests fail identically on clean HEAD — pre-existing, unrelated).

[Hmbown]

Both halves landed on v0914-chunk2:

• 4d70e050 — the UI no longer awaits the op channel on the input path (send_reserved_ capacity reservation off the render thread).
• e9de8a72 — Op::SendMessage carries TurnSpec in both variants (engine core::ops + serializable protocol::op), internally-tagged newtype so the wire shape is unchanged.

Verified via the commit diffs and the in-issue status trail.