daemon: batched dequeue per tick + SIGHUP config hot-reload

[mrap]

Motivation

Problem 1 — 40s ramp time

The daemon loop dequeued one spec per tick with a ~5s sleep. Starting from 0 workers with max_workers=8 took ~40s before all slots were filled. During that window specs sat idle in the queue.

Problem 2 — restart orphans in-flight workers

Bumping max_workers required boi daemon restart, which kills every running worker mid-task. This was the primary disruption path for live tuning.

What changed

spawns_per_tick (T2AD4 + T47CE)

• New spawns_per_tick: Option<u32> field in Config (default 4).
• Daemon loop now computes to_spawn = cap_remaining.min(spawns_per_tick) per tick and drains up to that many queue entries.
• Micro-jitter: 50–150ms randomized sleep between successive spawns within a tick smooths the Anthropic API cold-start burst.
• Each spawned worker logs its batch slot (batch slot 1/4) for easy telemetry correlation.
• boi config set spawns_per_tick N now works.

SIGHUP hot-reload (T7AFE)

• signal_hook::flag::register(SIGHUP, reload_flag) in the daemon.
• Before each tick: if flag set, config::try_load() → update wc.max_workers, wc.spawns_per_tick, wc.claude_bin. Parse failure = no-op + loud log.
• Only those three fields are live. Everything else stays frozen at startup. In-flight workers keep their original WorkerConfig.
• New boi daemon reload subcommand sends SIGHUP to ~/.boi/pids/daemon.pid. Typical path: boi config set max_workers 10 && boi daemon reload.

Mitigations

Risk	Mitigation
Anthropic API 429s from burst spawns	50–150ms per-spawn jitter; cap via spawns_per_tick
git worktree add index contention	Natural serialization: each spawn is sequential within the tick loop
Failure amplification	Each spawn's error breaks that slot only; loop continues
Bad SIGHUP config nuking live settings	try_load returns Err on parse failure; old values kept

Tests

cargo test --lib daemon_batch      # 8 tests — all passing
cargo test --lib daemon_hotreload  # 6 tests — all passing

Covered scenarios:

• Empty queue → 0 spawns
• 1 eligible, cap=4, tick=4 → 1 spawn
• 6 eligible, cap=4, tick=4 → 4 this tick (2 next tick)
• 6 eligible, cap=8, tick=4 → 4 this tick
• 4 eligible, cap=2, tick=4 → 2 this tick
• apply_reload updates only hot fields
• Bad config on SIGHUP → original wc retained, error logged
• Missing config on SIGHUP → defaults returned (no crash)

Cold-start behavior (smoke run)

Not yet run against live Anthropic API — no smoke data yet. Once this merges to main and a batch of 6+ specs are queued, observe batch slot logs and API 429 rate.

Docs

• docs/daemon.md: tick cadence, spawns_per_tick semantics, SIGHUP reload, what does/doesn't reload.
• README.md: boi daemon reload added to command list.
• SKILL.md: boi config set max_workers N && boi daemon reload is now the live-bump path.

---

Don't merge — Mike reviews.

🤖 Generated with Claude Code