fix interrupt latency and process-wake double turns

[jbj338033]

Summary

Two turn-lifecycle bugs plus the surrounding cleanup they exposed. Interrupts were ignored while a turn was parked in provider.stream(request).await (HTTP send + time-to-first-token), so pressing Esc did nothing until the request returned. And a watched background process re-woke the agent with a (process activity) turn even when the agent had already consumed that output/exit itself, because the read cursor and the watch cursor were tracked separately.

• interrupt: race provider.stream() open against token.cancelled() in run_round/compaction so cancellation drops the in-flight request immediately.
• process wake: unify read_cursor/watch_cursor into one seen_cursor and mark exit observed on read, so output already read via ProcessOutput never wakes a fresh turn; clarify the watch tool docs.
• cleanup surfaced along the way: coalesce process wakes through tokio::sync::Notify, share now_ms/truncate_chars, extract the triplicated mid-turn op pump, drop top_regime's constant sandbox tuple.

Verification

• cargo fmt --all --check — passed
• cargo clippy -p goat-agent -p goat-tool --all-targets -- -D warnings — passed
• cargo test -p goat-agent -p goat-tool — passed (113 + 31)