fix(search): resolve zero results bug and improve UX

[prosdev]

🐛 Problem

Search was indexing 566 documents but returning 0 results for all queries, making the semantic search feature completely non-functional.

Root Cause

LanceDB returns L2 distance (~1.0 for similar vectors), but our code incorrectly calculated:

score = 1 - distance = 1 - 1.0 = 0

All results scored 0 and were filtered out by the threshold.

---

✅ Solution

1. Fix Distance-to-Similarity Conversion

• Use exponential decay: score = e^(-distance²)
• Provides proper 0-1 range with good score distribution
• Exact matches now score 70-90%, semantic matches 25-60%

2. Fix CLI Metadata Access

• Changed metadata.file → metadata.path
• Prevents undefined errors in search output

3. Add Comprehensive Tests

• 11 new integration tests for search functionality
• Tests cover: stats, thresholds, limits, sorting, score ranges
• Protects against regression

4. Update Documentation

• Real working examples from testing
• Threshold recommendations (0.7=precise, 0.25=exploratory)
• Natural language query examples
• Actual output with real scores

---

🧪 Verification

Before:

$ dev search "coordinator" --threshold 0.7
✖ Found 0 result(s)

After:

$ dev search "coordinator" --threshold 0.3
1. CoordinatorLogger (42.6% match)
2. Coordinator - The Central Nervous System (42.4% match)
3. CoordinatorLogger.info (35.6% match)
✔ Found 3 result(s)

Test Results:

• ✅ 11/11 integration tests passing
• ✅ Natural language queries working ("how do agents communicate" → 51.9% match)
• ✅ Exact term matching ("RepositoryIndexer" → 85.7% match)
• ✅ Technical concepts ("vector embeddings" → 58.5% match)

---

📊 Impact

Search Quality Examples:

Query	Top Result	Score
RepositoryIndexer	RepositoryIndexer.index	85.7%
vector embeddings	EmbeddingProvider	58.5%
how do agents communicate	Message Architecture	51.9%
error handling	Handle Errors Gracefully	39.3%

Score Interpretation:

• 70-90%: Exact matches, highly relevant
• 40-60%: Strong semantic matches
• 25-40%: Related concepts, exploratory
• <20%: Weak matches

---

🔗 Commits (Atomic)

1. fix(search): Correct distance-to-similarity calculation
2. test(search): Add 11 comprehensive integration tests
3. chore: Update gitignore, remove tsx dependency
4. docs: Update READMEs with real examples

Each commit builds and tests independently.

---

🚀 Next Steps

• Fix explore similar command (searches filename as text, not content)
• Consider adding score calibration based on query length
• Add search result caching for frequently used queries

---

Dogfooded on dev-agent itself: All examples are real searches on this repository! 🐕