fix(semantic): strip code fences and canonicalize JSON before embedding

[babaliauskas]

Problem

semantic.cosine embedded raw output text, so presentation outweighed meaning. In a real run (r_20261011_triage_9fcdf0), 8/9 source outputs were ```json-fenced pretty-printed JSON and every target was compact JSON: every example scored 0.73–0.79 and was flagged SEMANTIC_REGRESSION. The one unfenced source scored highest (0.89) even though its category actually changed.

Fix

Before embedding, both outputs go through _meaning_text():

• drop a code fence wrapping the whole output (fences inside prose are part of the answer and kept);
• re-serialize a JSON object/array canonically (sorted keys, default separators);
• an empty fenced block counts as an empty output (avoids embedding "").

Always on, no config. Fence/validity checks stay with json_schema and the judge.

Evidence (live text-embedding-3-small, same run)

example	before	after
cap_c6ebf63e	0.733	0.966
cap_296ecd48	0.749	0.967
cap_8cd2b4c9	0.744	0.971
cap_c2cc1743 (real category change)	0.890 (highest)	0.908 (lowest)

0/9 flagged after the fix.

Tests / docs

• 5 new tests in TestCosineEvaluatorFormatNormalization (4 red before the fix; 1 guards that inner fences are kept).
• make ci: 2999 passed, 94.91% coverage; pre-commit run --all-files clean.
• DOCS.md, llms-full.txt, docs/evaluators.md, CHANGELOG.md (Unreleased → Fixed, notes scores rise on upgrade).

🤖 Generated with Claude Code