fix: robust extract_code_block (full-file block, nested fences)

[datj9]

What

ownbench/runner.py::extract_code_block had two defects that could silently zero a correct answer:

1. Last block wins. It returned blocks[-1]. When a model emits the complete fixed file first and then appends prose containing a small ```py snippet, the tiny (wrong) trailing block was returned instead of the full file.
2. Non-greedy capture + unanchored fence. The regex \``[^\n]\n(.?)```` used DOTALL with a closing fence that was not anchored to the start of a line. A ``` appearing inside file content (e.g. an indented fence in a docstring) ended the capture early, returning a truncated slice.

Fix

Replaced the single regex with a line-based scan:

• Opening fence: a line matching ^\s*```[^]*$` (leading whitespace allowed, optional info string), only when not already inside a block.
• Closing fence: a line matching ^```\s*$, i.e. three backticks at column 0 (optional trailing whitespace).
• Collects every top-level block (inner lines joined by \n, preserving the newline before the closing fence to match prior semantics).
• Returns the block with the greatest character length (max(blocks, key=len)), or None when there are none.

The full file outweighs a trailing snippet, and inner docstring fences are indented so they never match a column-0 closing fence. Signature/return type unchanged.

Tests added (tests/test_extract.py)

• test_trailing_snippet_ignored — full-file block followed by prose + a tiny ```py snippet; the full block must win.
• test_nested_fence_in_docstring — a block containing an indented fence inside a docstring must be returned whole, not truncated.
• test_regression — a single clean block is extracted exactly; text with no fence returns None.

Verification

• RED (before fix): the two defect tests fail.
• GREEN (after fix): full suite green — 18 passed / 0 failed.

Co-authored-by: CommandCodeBot noreply@commandcode.ai