fix: persist startRun and stopRunByError events for failed delegations

[FL4TLiN3]

Summary

• When a delegate's subprocess failed to start before the state machine initialized (e.g., MCP connection failure, npx not found in PATH), no events were persisted — the error was silently converted to a tool result text and returned to the parent expert
• The catch block in DelegationExecutor.executeSingleDelegation() now emits startRun + stopRunByError events via the parent's storeEvent and eventListener callbacks, ensuring failed delegations are recorded in the event store

Root cause

The child run() calls CoordinatorExecutor.execute() → SkillManager.fromExpert() → MCP adapter connect. If this fails, the error propagates before the state machine starts, so no startRun event is ever emitted and no run directory is created. The existing catch block only returned a text error to the parent — no structured failure record was persisted.

Evidence

Confirmed with real data from a create-expert job where 10 delegations (4× find-skill, 3× logic-engineer/tui-engineer/cli-engineer × 2) silently vanished. The parent's resumeFromStop checkpoint contained partialToolResults with "Delegation to X failed: Executable not found in $PATH: \"npx\"" — but no run directories or events existed for these delegations.

Test plan

• Existing delegation-executor.test.ts passes (16 tests)
• Typecheck passes for @perstack/runtime
• Verify with a real job that failed delegations now produce run directories with startRun + stopRunByError event files

🤖 Generated with Claude Code