fix(classifier): prefer feature intent for ambiguous prompts

[ozymandiashh]

Summary

CodeBurn was over-counting normal feature work as Debugging because the classifier checked broad debug keywords before feature intent. Words like error, issue, and fix are not always bug work: users often ask agents to add error handling, create issue trackers, or adjust layout for a new feature.

This PR narrows that behavior for the specific feature-style contexts called out in #196. It keeps explicit bugs, failures, crashes, status-code errors, and broken behavior in Debugging, but lets feature construction win for ambiguous debug-like phrases.

What changed

• Add a small feature-over-debug guard for error handling, issue tracker, and new feature contexts.
• Apply the guard to edit-tool coding turns and conversation-only turns.
• Preserve Debugging classification for explicit bug/failure prompts such as fix the login bug or fix the 500 error in the new feature.
• Add regression tests for the examples discussed in Counting does not work at all #196.
• Add an Unreleased changelog note.

Partially addresses #196.

Why this is safe

The change is limited to activity classification. It does not touch provider discovery, token parsing, pricing, session aggregation, or the separate provider-attribution/Zed gaps discussed in #196. Explicit debug signals still win when the prompt clearly describes a bug or failure.

Validation

• npx tsc --noEmit
• npx vitest run tests/classifier.test.ts
• npx vitest run
• npm run build