fix(runtime): respect max concurrent tasks in ExecutionEngine::execute_parallel()

[SH20RAJ]

Summary

Fixes #1505 - ExecutionEngine::execute_parallel() was spawning unbounded tokio tasks, completely ignoring the RuntimeConfig::max_concurrent_tasks setting, leading to potential resource exhaustion.

Problem

The execute_parallel() method spawned one tokio::spawn per input with zero concurrency limits. A caller passing 1,000 inputs would spawn 1,000 concurrent tasks simultaneously:

• Memory exhaustion
• File descriptor limit exceeded
• LLM API rate limit abuse
• Cascading system failures

The RuntimeConfig::max_concurrent_tasks setting was effectively dead code—users could set it to 5 through AgentBuilder but the executor would ignore it.

Solution

Implement Semaphore-based back-pressure:

• Acquire a permit before spawning each task
• Hold permit for task lifetime (prevents task from counting against limit)
• Respects concurrency constraints (default: 10 max concurrent tasks)
• Tasks queue naturally in executor; no busy-waiting

Impact

• Maximum concurrent tasks is now enforced
• No unbounded memory growth
• Graceful queueing under high load
• API rate limits protected
• System stability guaranteed

Testing

Tested with high-volume input (1000+ tasks) to verify:

• Only N concurrent tasks spawn at a time
• Memory remains bounded
• No tasks are dropped

GSOC 2026 Contribution

Critical bug fix for production stability. Addresses priority P1 issue #1505.

Cc: @Mustafa11300 @rahulkr182