Summary

• DagScheduler::record_spawn_failure() now detects transient concurrency-limit rejections and reverts the task to TaskStatus::Ready instead of Failed
• This prevents spurious graph failure cascades when SubAgentManager refuses a spawn because all concurrency slots are occupied (e.g., max_concurrent=1)
• Permanent spawn failures (invalid agent definition, config error) continue to mark the task Failed as before
• 2 unit tests added covering both error string variants produced by SubAgentManager

Root cause

DagScheduler.tick() optimistically marks tasks as Running before calling SubAgentManager::spawn(). When the manager rejects the spawn with "concurrency limit N reached", the old code called record_spawn_failure() which unconditionally set the task to Failed and cascaded failure to all dependents — aborting the entire plan.

Test plan

• cargo +nightly fmt --check — pass
• cargo clippy --workspace --features full -- -D warnings — pass
• cargo nextest run --workspace --features full --lib --bins — 4986 passed (+2 new tests), 11 skipped

Follow-up

• Typed SubAgentError::ConcurrencyLimit variant to replace string matching (filed separately)
• Additional edge-case tests: multi-task deferral, max_concurrent=0, deadlock detection

Closes #1513