feat(search): implement normalized weighted score fusion for hybrid retrieval

[spelech]

Summary

This PR replaces Qdrant server-side Reciprocal Rank Fusion (Fusion.RRF) with Normalized Linear Weighted Fusion ($\alpha \cdot \text{Dense} + (1 - \alpha) \cdot \text{Sparse}$, with default $\alpha = 0.7$).

Problem

Previously, Qdrant hybrid search used Reciprocal Rank Fusion (Fusion.RRF). Because RRF calculates scores strictly from reciprocal rank fractions ($1 / (1 + \text{rank})$), top-ranked results scored in the 0.20 – 0.50 range (e.g. RRF Score: 0.3333). In the UI and MCP tools, this appeared counter-intuitively low (as if relevance was only 33%), and diverged from single-vector cosine scoring in pgvector and Chroma (0.70 – 0.95). Furthermore, candidate prefetch was capped at limit * 2, which caused cross-modality cutoff mismatches.

Solution

1. Multi-Stream Batch Retrieval:
   • Dispatches both dense and sparse query requests concurrently via client.query_batch_points in a single network roundtrip.
   • Expands candidate pool depth to $\max(\text{limit} \times 5, 50)$ so high-ranking items in either modality have sufficient depth to intersect.
2. Score Normalization & Linear Weighted Fusion:
   • Dense cosine similarity is clamped to [0.0, 1.0].
   • Sparse BM25 scores are normalized relative to the maximum observed BM25 score in the retrieved candidate batch ($S_{\text{sparse}} / \max(S_{\text{sparse_max}}, 1.0)$).
   • Scores are linearly combined with weight $\alpha = 0.7$, configurable via HYBRID_DENSE_WEIGHT environment variable.
   • All output scores strictly fall within [0.0, 1.0].
   • Seamless fallback to pure dense cosine search when sparse tokens are empty or unavailable.
3. UI & MCP Output Formatting:
   • Search Inspector UI badge updated to Score: XX.X% (X.XXXX) (e.g., Score: 80.5% (0.8049)).
   • MCP search tools (search_code, search_docs) updated to Relevance Score: X.XXXX (XX.X%).
   • Updated retrieval strategy label in Overview tab to Dense + BM25 Normalized Weighted Fusion.
4. Rebuilt Production Bundle & Synced Catalog:
   • Rebuilt frontend production distribution in frontend/dist.
   • Synchronized test matrix and catalog in REQUIREMENTS.md (934 automated tests baseline).

Verification

• Pytest: 496 passed (100%).
• Vitest: 289 passed across 27 test files (100%).
• Requirements Sync: test_requirements_sync.py passed (100%).