feat(reasoning): add reasoning effort configuration to providers

[jrob5756]

Summary

Adds an optional unified reasoning.effort field that lets users dial up
model "thinking" / reasoning at both the workflow level
(runtime.default_reasoning_effort) and per-agent (reasoning.effort).
Per-agent value overrides the runtime default.

runtime:
  provider: copilot
  default_reasoning_effort: medium

agents:
  - name: explainer
    model: gpt-5
    prompt: \"Explain this algorithm.\"
    # inherits runtime default 'medium'

  - name: architect
    model: gpt-5
    reasoning:
      effort: high      # override
    prompt: \"Design a system architecture.\"

Allowed values: `low`, `medium`, `high`, `xhigh`.

Per-provider translation

Effort	Copilot	Claude (`thinking.budget_tokens`)
low	`reasoning_effort="low"`	2048
medium	`reasoning_effort="medium"`	8192
high	`reasoning_effort="high"`	16384
xhigh	`reasoning_effort="xhigh"`	32768

Copilot

• Forwards `reasoning_effort` to `CopilotClient.create_session`.
• Validates against the model's `supportedReasoningEfforts` (from
  `list_models()`) and raises `ValidationError` with a clear message
  when the model does not support the requested effort.
• Skipped in mock-handler mode.
• Retry loops fixed to not swallow `ValidationError`.

Claude

• Enables extended thinking via
  `messages.create(thinking={"type":"enabled","budget_tokens":N})`.
• Auto-coerces `temperature=1.0` (Anthropic API requirement) and bumps
  `max_tokens` to fit `budget+4096` (capped at 64000 for thinking-
  enabled requests).
• Validates against thinking-capable model prefixes (`claude-3-7-`,
  `claude-opus-4`, `claude-sonnet-4`, `claude-haiku-4`) and raises
  `ValidationError` for non-thinking models.
• Surfaces thinking content blocks via the existing `agent_reasoning`
  event (provider parity with Copilot's `assistant.reasoning`).

Provider parity

Both providers:

• Accept the same `reasoning.effort` field with identical semantics.
• Validate model support and raise a consistent `ValidationError` when
  unsupported.
• Emit `agent_reasoning` events from any reasoning content the model
  produces, so the dashboard, JSONL logger, and console subscriber
  render it consistently.

Schema

• `AgentDef.reasoning: ReasoningConfig | None` (new `ReasoningConfig`
  with single `effort` field).
• `RuntimeConfig.default_reasoning_effort: Literal[...] | None`.
• Forbidden on `script`, `human_gate`, and sub-workflow agent types
  (parity with how `model` and `retry` are handled).

Files changed

• `src/conductor/config/schema.py` — new `ReasoningConfig`,
  `AgentDef.reasoning`, `RuntimeConfig.default_reasoning_effort`,
  forbid rules.
• `src/conductor/providers/reasoning.py` (new) — shared effort→budget
  mapping, Claude thinking-model prefix allow-list, agent/runtime
  resolver.
• `src/conductor/providers/copilot.py` — forward `reasoning_effort`,
  validate against `supportedReasoningEfforts`, retry loop fix.
• `src/conductor/providers/claude.py` — forward `thinking` kwarg at
  every `messages.create` site, auto-coerce temperature/max_tokens,
  emit `agent_reasoning` from thinking blocks.
• `src/conductor/providers/factory.py` — typed `ReasoningEffort`
  forwarding to both providers.
• `examples/reasoning-effort.yaml` (new) — demonstrates inheritance
  and per-agent override.
• `AGENTS.md` — key pattern + provider parity bullet.

Tests

45 new tests:

• 21 schema cases (parametrized): accept/reject effort values, runtime
  default, per-type forbid rules.
• 7 Copilot cases: `reasoning_effort` forwarded; runtime default and
  per-agent precedence; key absent when unset; `ValidationError` on
  unsupported effort; mock-mode skip.
• 13 Claude cases: `thinking` kwarg shape; effort→budget mapping;
  `temperature` coerced to 1.0; `max_tokens` bumped; `ValidationError`
  on non-thinking model; runtime/per-agent precedence; thinking blocks
  emit `agent_reasoning`; key absent when unset.
• 4 e2e + factory wiring cases.

Verification

• `make lint` ✅
• `make typecheck` ✅ (only a pre-existing unrelated warning in
  `dialog_evaluator.py` remains)
• `make validate-examples` ✅ (all 14 examples)
• `make test` ✅ — 2330 passed, 9 skipped

Out of scope

• Anthropic `display: summarized | omitted` flag.
• Anthropic `adaptive` thinking type.
• Mid-workflow dynamic effort adjustment.

🤖 Generated with Copilot CLI