refactor(epic-1973): Extract zeph-subagent crate from zeph-core (Phase 1f)

[bug-ops]

Summary

• Extract 9,036 LOC subagent management module from zeph-core into standalone zeph-subagent crate
• zeph-core reduced from 64,851 to 55,860 LOC (-13.9%)
• No circular dependencies: zeph-subagent has zero imports from zeph-core

New crate: crates/zeph-subagent

11 modules: command, def, error, filter, grants, hooks, manager, memory, resolve, state, transcript

Key design decisions

spawn_for_task() callback refactor — instead of taking TaskId + mpsc::Sender<TaskEvent> (which would require zeph-subagent → zeph-core), the method now accepts a generic completion callback:

pub fn spawn_for_task<F>(&mut self, ..., on_done: F) -> Result<String, SubAgentError>
where F: FnOnce(String, Result<String, SubAgentError>) + Send + 'static

The caller (agent/mod.rs) constructs the closure that sends TaskEvent to the scheduler. This breaks the orchestration dependency at the API boundary.

Re-export shim — zeph-core/src/lib.rs replaces pub mod subagent; with:

pub mod subagent {
    pub use zeph_subagent::*;
    pub use zeph_subagent::{command, def, error, filter, grants, hooks, manager, memory, resolve, state, transcript};
}

All crate::subagent::* paths in orchestration, agent, and tests work unchanged.

SubAgentHandle fields — widened from pub(crate) to pub to allow test harnesses in zeph-core to construct handles via struct literals.

Test results

• 5,917 tests passing (299 new in zeph-subagent), 0 failures
• cargo +nightly fmt --check: clean
• cargo clippy --workspace --features full -- -D warnings: clean

Breaking changes

None. All re-exports in zeph-core maintain backward compatibility.

Next phase

Phase 1g will extract zeph-orchestration. spawn_for_task no longer depends on orchestration types, so the circular dependency between orchestration and subagent is resolved.

Spec: .local/specs/003-epic-1973-phase-1f/spec.md
Epic: #1973