feat(delegate): add task-tier profiles with per-task reasoning_effort and pool validation

[MestreY0d4-Uninter]

Summary

Unified delegation tier system combining named task profiles (model/routing/reasoning/iterations per tier) with optional model pool validation. Addresses open issues #3719, #6306 and supersedes the tier concept from stale PR #5692.

What this adds

5 named tiers — light, heavy, review, planning, research — each configuring model, provider, reasoning_effort, and max_iterations:

delegation:
  default_tier: heavy
  tiers:
    light:    {model: gpt-5.4-mini, reasoning_effort: low,  max_iterations: 25}
    heavy:    {model: gpt-5.4,      reasoning_effort: medium, max_iterations: 50}
    review:   {model: gpt-5.4,      reasoning_effort: xhigh, max_iterations: 60}
    planning: {model: xiaomi/mimo-v2-pro, provider: nous, reasoning_effort: high, max_iterations: 60}
    research: {model: gpt-5.4,      reasoning_effort: high, max_iterations: 60}

Reasoning floor guardrails prevent silent degradation:

• heavy/research → reasoning ≥ medium
• planning/review → reasoning ≥ high

Per-task tier in batch mode:

delegate_task(tasks=[
    {"goal": "Quick lookup", "tier": "light"},    # cheap/fast
    {"goal": "Code review", "tier": "review"},    # high reasoning
])

Optional model pool for validation (inspired by PR #5229):

delegation:
  pool:
    - model: gpt-5.4-mini, strengths: quick lookups
    - model: gpt-5.4, provider: openai-codex, strengths: coding, debugging

Resolution order

task.tier → top-level tier → default_tier → flat config → parent inherit

Files changed

File	Change
tools/delegate_tool.py	resolve_tier_config(), _validate_pool_model(), _build_pool_description(), override_reasoning_effort in _build_child_agent(), per-task tier in batch loop, schema updates
cli-config.yaml.example	Documented tiers, pool, reasoning_effort with examples
tests/tools/test_delegate_tiers.py	56 unit tests: tier resolution, reasoning floors, pool validation, schema, integration, comparative, backward compat, edge cases
tests/tools/test_delegate_tiers_real.py	7 real integration tests with cost comparison

Test results

• 128 total delegate tests passing (67 existing + 56 new tier tests + 5 toolset scope)
• 7 real integration tests passing (comparative cost analysis, schema, pool, backward compat)
• All existing tests pass unchanged (backward compatible)

Related issues

• Closes Feature: Allow per-task model selection in delegate_task #6306 (per-task model selection)
• Related to Feature: Per-call model/provider override on delegate_task #3719 / PR feat(tools): add delegation model pool with per-call model/provider override #5229 (per-call model/provider override)
• Extracted from refreshed audit of stale PR fix(credential-pool): harden multi-account pool — health scoring, dynamic cooldown, leases, mark_success #5692

Breaking changes

None. Without tiers configured, behavior is identical to before.

[MestreY0d4-Uninter]

Closing — needs more real-world testing and calibration before submission. Will reopen with comprehensive data.