feat(engine): stop/resume workflow control with provider-level idle detection

[jrob5756]

Summary

Stop/Resume workflow control for the web dashboard, provider-level idle detection with interrupt support, and dashboard state resilience across page refreshes.

Changes

Provider-level idle detection + interrupt (Copilot)

• Merged interrupt support INTO the idle detection loop so both work simultaneously
• Previously, having an interrupt_signal completely bypassed idle recovery, causing sessions to hang forever
• Now the provider monitors for user interrupts AND sends recovery prompts after 90s of silence, while respecting active work (events flowing = not idle)
• Lowered idle timeout from 300s to 90s for faster stall recovery
• Recovery prompt ("please continue") is harmless to active work — model ignores it if already processing

Provider interrupt during API calls (Claude)

• Race interrupt signal against both _execute_api_call and _execute_with_parse_recovery for provider parity
• Copy working_messages in _request_partial_output to avoid mutating caller's history

Web dashboard Stop → Resume

• Stop (POST /api/stop): Aborts the current agent via interrupt event (not kill workflow)
• Resume (POST /api/resume): Re-executes the paused agent
• Kill (POST /api/kill): Hard-stop with checkpoint saved for CLI resume
• Separate kill_event from _stop_event to prevent permanent poisoning across pause cycles
• disconnect_event for auto-resume when browser closes during pause (prevents hanging forever)
• agent_paused / agent_resumed events for dashboard state tracking

Dashboard state resilience (page refresh)

• workflow_started handler uses event timestamp instead of Date.now() — elapsed timer survives refresh
• agent_started / script_started store startedAt from event timestamp — node elapsed timers survive refresh
• replayState sets lastEventTime from last replayed event — idle timer works after refresh
• useLiveElapsed hook uses stored startedAt instead of Date.now() on mount

Web mode interrupt support

• Always create interrupt_event in --web/--web-bg mode so dashboard Stop button reaches the engine
• Share interrupt_event with dashboard via set_interrupt_event()

Other

• Remove revive infrastructure (Ctrl+R, /api/revive, revive watcher) — replaced by idle detection + Stop/Resume
• Remove dead _wait_with_interrupt method (merged into idle detection)
• Hoist json import to module level in workflow engine
• Silently consume stale interrupt in web mode between-agent checks
• Bump flaky CLI import time test threshold from 0.5s to 1.0s
• Header: useEffect for button state resets, error logging in catch blocks, killing state
• Upgrade Copilot SDK dependency to >=0.2.0

Test results

• 1760 passed, 9 skipped
• Lint clean, typecheck clean
• Frontend builds successfully