fix(mcp): make interactive tool calls transport-drop-safe (supersede, reap, keep-alive)

[juacker]

Problem

When the Claude Code CLI drops the local MCP transport mid-call (transport dropped mid-call; response for tool <name> was lost) during an interactive tool call — ask_user, an approval-gated bash_exec, or fs_request_grant — the in-flight wait was orphaned:

• rmcp's stateful streamable-HTTP server runs call_tool in a detached session worker (spawn_session_worker) that survives the dropped connection, so the abandoned-wait RAII guards (which fire on future drop) never ran, despite their doc comments claiming they handled transport drops.
• The pending registry entry and its approval/path-grant card lingered for up to 55 minutes (CLI_INTERACTIVE_WAIT_TIMEOUT); the run wasn't stopped.
• A stale ask_user entry also kept session_has_pending_ask true, suppressing mid-run input delivery.
• If the model re-asked, the user saw a duplicate card stacked next to a stale, unanswerable one.

Fix (three layers, root-cause first)

1. Supersede on re-ask — registering a new interactive request takes any stale pending entry for the same key (bash: run + command; path grant: run + path + access; ask_user: session, matching the FE's single-panel model), emits resolved/attention so the stale card is replaced in place the moment the fresh one appears, and wakes the orphaned wait with a channel-closed error that is now treated as superseded (guard disarmed — never cancels the live run that's waiting on the new request).
2. Run-end reaping — BindingGuard now owns a run-scope CancellationToken, cancelled on drop; execute_bound_tool races it. Any tool future still in flight when the run ends (only possible for orphans — a live CLI waits for its tool results) is dropped, which finally fires the existing cleanup guards exactly as they were documented to.
3. Keep-alive restored — sse_keep_alive is left at rmcp's default Some(15s) instead of None. The None was carried over from the pre-1.7 struct literal, not a decision; an idle unpinged response stream during a long human wait is precisely what rots into a transport drop.

Consequential fix: workspace_delete now explicitly cancels the runs whose pending approvals it purges (purge_workspace returns the awaiting run ids), since a closed approval channel no longer implies run cancellation.

Prompt change (first commit, amended by the second)

The scoped Interactive Tool Reliability section (emitted only when an interactive tool is present) still tells the model a dropped interactive call has an UNKNOWN outcome and to re-issue the same call once — but the "tell the user they can dismiss the stale card" caveat is gone: superseding replaces the card automatically.

Tests

• BindingGuard drop cancels the run scope + unbinds without touching the run's cancel token; the reap arm drops the racing future (RAII guards fire).
• take_superseded matches only same run + command / run + path + access; channel-close as supersede signal; counts maintained.
• ask_user::take_for_session scoping + superseded submit_answer rejection.
• Prompt tests assert the dismiss caveat is gone.

cargo test --lib: 611 passed. cargo fmt --check, cargo clippy --lib --tests, tsc --noEmit clean.