tech-debt: handshake TTL with acp_handshake_stuck event (post-#4761)

[OneStepAt4time]

Context

Follow-up to PR #4761 (race-regression fix for #4760). Argus surfaced this as a deferred hardening item: "Handshake TTL with acp_handshake_stuck event."

If a background handshake hangs (network blackhole, runtime deadlock, partially-initialized state), pendingHandshakes retains the entry indefinitely. Downstream sendPrompt awaits the dead ready until the process itself times out. No signal to operators that the handshake is stuck.

Acceptance criteria

• Configurable per-handshake TTL (default e.g. 60s)
• When TTL fires before ready settles: emit acp_handshake_stuck event with sessionId + backendRunId + age + last-known-state
• Remove the Map entry on TTL fire (so subsequent sendPrompt falls through to the canonical error path, not the dead await)
• Test: mock handshake promise that never settles → after TTL, acp_handshake_stuck fires once + Map entry removed
• Test: handshake that settles BEFORE TTL → no event fires, normal flow

Out of scope

• Max-size + LRU eviction (separate issue)
• ReadonlyMap type-level tightening (separate issue)

Lane

Backend (Hephaestus). Tracking follow-up to PR #4761 / issue #4760. File: P2 tech-debt; do not block PR #4761 merge.

— Athena 🦉 per Argus 9-gate review surface