fix(router): route synthetic wake busy through the session manager

[yicheng47]

Fixes #385.

Root cause

Two independent "is this runner busy" stores could diverge: the router's synthesize_wake_busy appended a synthetic runner_status busy directly to the log without updating session.activity — the dedup key the forwarder consults. When the agent's first output byte landed inside the suppressed-busy window, the forwarder's busy was swallowed with activity still Idle, so the end-of-turn idle was deduped as a no-op and the rail stayed busy forever. A stuck handle is also permanently ineligible for the reconciliation nudge, so this could silently stall missions.

Changes

• synthesize_wake_busy now routes its append through the SessionManager, which writes the event and updates session.activity under one session lock — the two stores can no longer drift, and the existing dedup works as intended. Wakes at a dead slot no longer append an uncleareable busy (early return instead).
• try_append_runner_status retries on flock contention (8 attempts × 5 ms, bounded < 50 ms) instead of dropping — status events are load-bearing now that the reconciliation gate reads them.
• Regression tests: both stores updated by the synthetic busy with the final idle emitted; the suppressed-busy → output → quiet sequence still yields a logged idle; bounded retry recovers a contended append; retry budget stays bounded under persistent contention.

Checks

• cargo test --workspace (540 runner + 8 CLI + 16 roundtrip + 22 runner-core), cargo clippy, cargo fmt --check — all green locally.
• Crew-reviewed on the working-tree diff (mission 01KZMYBZPR…): no must-fix findings; both reviewer nice-to-haves folded in.

🤖 Generated with Claude Code