fix: make ParallelExecutor the single timeout authority (#760)

[SBALAVIGNESH123]

What

Removes the hardcoded 600s MCP timeout from AgentStrategy.execute_mcp_tool and establishes PE as the single source of truth for all tool deadlines.

Why

After #755 introduced parallel execution, timeout enforcement ended up scattered across two layers:

1. PE sets a Process.send_after deadline per tool (from ToolExecution.timeout_ms)
2. AgentStrategy.execute_mcp_tool has its own receive..after 600_000

These race each other. If PE's deadline fires first, it kills the task — but the inline after might also fire in edge cases. Worse, the 600s value is hardcoded and ignores the tool definition entirely, so interactive tools that should timeout at 120s actually wait 600s.

The fix is simple: PE already has everything it needs. The strategy just needs to get out of the way.

Changes

agent_strategy.ex — the actual fix

• Replaced receive..after 600_000 with bare receive (no after)
• PE kills the process on deadline, Registry auto-cleans up the key — no manual timeout needed
• Added comment explaining the contract so nobody re-adds a timeout here

runtime.ex — default timeout to infinity

• await_tool_result default changed from 600_000 to :infinity
• Same reasoning: PE owns the deadline, not the callee
• Explicit timeout: opt still works for tests and escape hatches

default_execution_strategy.ex — dead code removal

• Removed the {:error, :timeout} branch that is now unreachable

parallel_executor.ex + execution_strategy.ex — docs

• Added Timeout Ownership Contract section documenting PE as the single authority
• Documents the end-to-end flow: tool definition -> ToolExecution struct -> PE deadline -> on_timeout callback

timeout_ownership_test.exs — new tests

• PE kills a process stuck in bare receive
• on_timeout callback fires correctly
• Per-tool timeout isolation (fast tool times out, slow tool succeeds)
• Registry key auto-cleanup on process kill
• :pause_agent works with bare receive

How to verify

mix test apps/swarm_ai/test/swarm_ai/timeout_ownership_test.exs
mix test apps/swarm_ai/test/swarm_ai/parallel_executor_test.exs
mix test apps/swarm_ai/test/swarm_ai/runtime/tool_delivery_test.exs

All existing PE and delivery tests pass unchanged — they already use explicit timeout: values.

Answers to the questions from #760

Should PE know about tool types at all?

No. PE is type-agnostic — it reads timeout_ms and on_timeout_policy from the ToolExecution struct. The strategy decides what those values are.

Should the 60s timeout cover the full sub-agent tree?

For now, yes — sub-agents run inside the parent's Sync task, so PE's deadline covers the whole tree. Decoupling sub-agent timeouts is a bigger change and should be separate.

Is Task.async (linked) the right primitive?

Not relevant anymore — PE uses Task.Supervisor.async_nolink, which is correct.

Could we give each tool a timeout hint as metadata?

SwarmAi.Tool.timeout_ms already does this. The problem was that AgentStrategy was not using it. Now it does not need to — PE reads it from ToolExecution.timeout_ms and enforces it.

Closes #760

---

[SBALAVIGNESH123]

Addressed the two new findings from Devin Review:

1. ToolExecutor.register_mcp_tool references removed ToolCallRegistry — This is pre-existing from PR #902 (ExecutionStrategy migration). ToolExecutor is legacy code that's only reachable through sub-agent paths now. The primary production path routes through AgentStrategy which uses the Runtime's ToolRegistry instead. Migrating sub-agent paths off ToolExecutor is planned as follow-up work — not in scope for this timeout-ownership PR.

2. submit_to_runtime exceeds 70-line limit — Fixed in ca363fe. Extracted the tool_executor closure into a separate build_tool_executor/1 helper. submit_to_runtime is now ~55 lines, well under the limit.