fix(#21): Enforce strict parsing of LLM method output

[AK11105]

Summary

Fixes issue #21: _parse_methods conflates load() and predict() into one block.

Problem

The original implementation used lookahead regex ((?=\ndef |\Z)) to split methods from LLM output. This failed in three scenarios:

1. Class-wrapped output: When the LLM wraps methods in a class despite the system prompt, both load_match and predict_match capture the entire class body, causing both methods to be identical.
2. Missing blank lines: Without a blank line between methods, the lookahead \ndef doesn't match properly.
3. Trailing text: Text after predict() with no subsequent def gets included in predict_body.

The bug manifests at runtime: self._model is never set in the _GeneratedModel class, raising AttributeError on first prediction.

Solution: Strict Parsing

Replace lookahead regex with explicit def boundary splitting AND enforce the system prompt contract:

blocks = re.split(r"(?=^def )", raw, flags=re.MULTILINE)
# Reject any non-method content (class wrappers, trailing text)
# Enforce: "Return ONLY the two method bodies as plain Python"

This approach:

• Splits on start-of-line def — Rejects indented methods (class wrappers)
• Validates no trailing content — Each block must be a valid method definition
• Raises clear errors — Non-compliant output triggers LLM retry via fix()
• Prevents silent failures — Better to fail validation than silently corrupt self._model

Why Strict?

If we accept edge cases (class wrappers, trailing text), the LLM learns it can ignore the system prompt. Strict parsing enforces the contract, preventing more bugs downstream.

Testing

• test_parse_methods_no_blank_line_between ✅ Accepts (valid: missing blank line)
• test_parse_methods_class_wrapped_raises ❌ Rejects (violates: no class wrapper)
• test_parse_methods_trailing_text_raises ❌ Rejects (violates: ONLY two methods)

All existing tests continue to pass.

[AK11105]

Copilot reeks