Add self-improvement proposal mechanism

[sparklkt]

What changed

This PR adds a non-autonomous self-improvement proposal layer on top of the evaluation and growth-log system.

• adds ImprovementProposal as a structured proposal model
• adds proposal generation from evaluation failures and growth-log memory signals
• supports proposal types for scoring rule adjustment, factor template addition, collision lane adjustment, test case addition, documentation update, and config tuning
• adds CLI support for innerbrain propose-improvements
• keeps all proposals human-gated: requires_human_approval = true and safe_to_auto_apply = false
• adds tests for repeated failures, human-approval enforcement, and low-risk documentation proposals that still do not auto-apply

Why it changed

After adding evaluation and benchmark scenarios, the next useful step is not autonomous self-modification. It is structured self-critique: the system should be able to identify where rules, factors, collision lanes, tests, or documentation likely need revision, then hand those ideas to a human reviewer.

Validation

• /Users/qianjinus/Documents/类脑/innerbrain-factor/.venv/bin/pytest -> 31 passed in 0.19s

Safety boundary

• remains offline and rule-based
• no LLM integration
• no network participation in reasoning
• no real external actions
• proposals do not modify the system automatically
• proposals do not bypass human approval, including low-risk documentation suggestions