feat(delegation): support per-call child model overrides

[JiraiyaETH]

Summary

• add per-call delegate_task model overrides via model={provider, model}
• add top-level and per-task reasoning_effort overrides
• resolve delegation credentials per task so batch delegation can fan out across configured providers/models
• forward the new public schema fields through AIAgent._dispatch_delegate_task

Validation

• python -m py_compile tools/delegate_tool.py run_agent.py tests/tools/test_delegate.py
• python -m pytest tests/tools/test_delegate.py::TestDelegateRequirements::test_schema_valid tests/tools/test_delegate.py::TestDelegationProviderIntegration::test_per_call_model_override_reaches_child_agent tests/tools/test_delegate.py::TestDelegationProviderIntegration::test_per_call_model_only_override_reuses_delegation_provider tests/tools/test_delegate.py::TestDelegationProviderIntegration::test_per_call_model_only_override_can_inherit_parent_provider tests/tools/test_delegate.py::TestDelegationProviderIntegration::test_batch_mode_can_mix_per_task_models tests/tools/test_delegate.py::TestDelegationProviderIntegration::test_single_task_reasoning_effort_reaches_child_agent tests/tools/test_delegate.py::TestDelegationReasoningEffort::test_per_call_reasoning_effort_overrides_config tests/tools/test_delegate.py::TestDispatchDelegateTask::test_agent_dispatch_forwards_model_and_reasoning -o 'addopts=' -q → 8 passed
• python -m pytest tests/tools/test_delegate.py -o 'addopts=' -q -k 'not heartbeat_does_not_trip_idle_stale_while_inside_tool' → 146 passed, 1 deselected
• python -m pytest tests/cli/test_cli_interrupt_subagent.py tests/run_agent/test_tool_executor_contextvar_propagation.py -o 'addopts=' -q → 6 passed
• direct provider smoke: openai-codex/gpt-5.4-mini → GPT54_MINI_DIRECT_OK
• e2e delegation smoke: DeepSeek child override → DEEPSEEK_PARENT_OVERRIDE_OK
• e2e mixed batch smoke produced both task completion markers: BATCH_SYNC_DEEPSEEK_OK, BATCH_SYNC_GPT54_OK

Known unrelated test failure

tests/tools/test_delegate.py::TestDelegateHeartbeat::test_heartbeat_does_not_trip_idle_stale_while_inside_tool is currently failing on origin/main as well as this branch. Verified in a separate disposable verifier worktree reset to origin/main:

HEAD is now at 87ab37338 test(url-safety): cover IPv6 scope-ID strip + fail-closed in URL guards
FAILED tests/tools/test_delegate.py::TestDelegateHeartbeat::test_heartbeat_does_not_trip_idle_stale_while_inside_tool

This PR leaves that heartbeat test untouched.

[alt-glitch]

Related: #47295, #32711 (feature requests for per-call model routing in delegate_task), and the open competing PRs #48867, #50015, #17756, #12715 (and others) that implement the basic per-call/per-task model override.

This PR is a superset — it adds reasoning_effort overrides and per-task credential resolution for batch fan-out across configured providers on top of the model override. Linking the cluster so reviewers can consolidate on one approach.

[teknium1]

We do not want this

[davidrobertson]

We do not want this

@teknium1 Many users do, and I'd love to have a discussion around this. At the very least, we could gate the allowed behavior behind an opt-in config option, maybe something like `delegation.dynamic_subagents: true' (defaults to false).

There are benefits to allowing this, such as having higher model/reasoning effort planning/review subagents for projects with larger sets of dependencies or security implications.

Thoughts?