fix: Feature grouping and Java language support (#140, #143, #146)

[mmorris35]

Summary

Fixes the regression from the previous PR and the root cause of the Java/.txt issue.

#146 - Feature grouping over-corrects (REGRESSION)

• Root cause: groupRelatedFeatures only pushed groups when they had items (endpoints)
• Fix: Now pushes ALL groups - features without endpoints are preserved as separate phases
• Features 1-7 now appear correctly, with endpoints grouped under last feature before them

#140 - Plan scaffold generates .txt paths for Java

• Root cause: TemplateKey.language type only included 6 languages (python, typescript, javascript, go, rust, unknown)
• When detectLanguage returned "java", switch statement fell through to "unknown"
• getLanguageDefaults("unknown") returned .txt extension
• Fix: Added java, kotlin, csharp, ruby, swift, cpp, zig to TemplateKey.language
• Updated switch statement to handle all detected languages properly

#143 - Endpoints as separate phases

• Fixed by bug: Feature grouping over-corrects — 6 of 7 features dropped, only Leaderboard remains #146 fix - endpoints now grouped with preceding feature

Testing

• wrangler deploy --dry-run passes

Remaining Issues

• bug: CLAUDE.md and agent generators ignore detected language, emit Python/TypeScript defaults #139 - Executor/Verifier agents still need language-aware templates (CLAUDE.md was fixed)
• bug: Lessons learned not filtered by project type — irrelevant Graph API/Cargo/RocksDB lessons in Java plan #142 - Lesson filtering may need testing to verify it's working

Closes #140, #143, #146