chore(release): 1.4.3

[babaliauskas]

Bumps to 1.4.3 and cuts the CHANGELOG section for PR #53:

• Failure categories count each example once, not each scored row (TOOL_SELECTION_DRIFT was double-counted, once from the conformance row and once from the divergence row).
• The cached run narrative regenerates when the run's scores, analysis or decision change, so it can no longer read "FAIL" beside a PASS verdict after a re-analyze.

Docs: #53 already updated DOCS.md, llms-full.txt and docs/configuration.md. This PR only bumps the version lines.

The pre-push CI mirror passed (ruff, format, mypy --strict, pytest). After merge, tag the merge commit v1.4.3 and push the tag to publish.

🤖 Generated with Claude Code