Raise default agent.max_concurrent from 4 to 32

[pufit]

Why

engine.run() holds a global asyncio.Semaphore(agent.max_concurrent) for the entire duration of an agent turn. With the default of 4, a handful of concurrent long-running sessions (e.g. several deep investigations, Claude + Codex side by side, plus cron jobs — all competing for the same pool) saturate every slot. Any further message — to an existing session, the first turn of a brand-new session, or a scheduled cron turn — parks silently at the semaphore: no session_running broadcast, no message persistence, no title, no streaming. To the user the instance looks unwritable until a long turn finishes and frees a slot (observed queue delays of 8+ minutes).

Turns are almost entirely I/O-bound (awaiting the backend CLI subprocess), so a much higher ceiling is safe; 32 keeps the semaphore as a runaway backstop rather than a throughput limiter.

Changes

• nerve/config.py: AgentConfig.max_concurrent default 4 → 32 (dataclass default + loader fallback)
• nerve/bootstrap.py: setup-wizard generated config → 32
• config.example.yaml, docs/config.md: documented default → 32

Existing configs with an explicit max_concurrent are unaffected.

Testing

• Full suite: 1588 passed, 2 skipped.

Follow-up (not in this PR)

The silent-queueing UX is a separate issue: messages waiting on the semaphore aren't persisted or acknowledged, so users can't tell "queued" from "broken" (and queued messages are lost on restart). Worth persisting the user message and broadcasting a queued state before semaphore acquisition.

Generated by Nerve