P2: Add thinking level as estimation parameter

[haoranc]

Summary

All calibration data (83+ validated dispatches) was collected with elevated thinking levels:

• Claude Code: high thinking
• Codex: extra high thinking

This is not captured anywhere in the estimation model. Thinking level affects:

• Cost per turn: High thinking uses significantly more tokens per turn (~2x for Opus)
• Turns needed: Higher thinking = fewer iterations (gets it right first time)
• Duration per turn: Higher thinking = longer turns (~2-3 min vs ~1-1.5 min at default)
• Quality: Consistently Q4-Q5 scores may partly be because of elevated thinking

Proposed Changes

1. New --thinking-level CLI parameter

ae estimate --thinking-level high "Add user authentication"
ae estimate --thinking-level default "Add user authentication"  # lower cost, more turns

Values: default, high, extra-high

2. Thinking level modifier in PERT engine

Thinking Level	Cost/Turn Modifier	Turns Modifier	Net Duration
Default	1.0x	1.0x	baseline
High	1.5-2.0x	0.7-0.8x	~1.2-1.6x cost, ~0.8x duration
Extra High	2.0-2.5x	0.6-0.7x	~1.5-1.8x cost, ~0.7x duration

3. Document in calibration data

Note which thinking level each calibration entry was collected at. Current entries are all high/extra-high.

4. Default behavior

If --thinking-level is not specified, default to high (matches calibration baseline). Flag in output: "Estimated at [thinking level] — calibration baseline."

Context

The Iris /estimate skill has been updated to document this gap. The agents.md memory file now records default thinking levels per agent. This issue tracks the product-side implementation.

Priority

P2 — important for accuracy but existing estimates are implicitly correct for high/extra-high users (which is our primary use case). Becomes P1 if/when external users with different thinking configurations report estimate drift.