feat: Try with reranker

[GoodbyePlanet]

What was built:

• eval/ — a new evaluation framework with a query dataset (14 queries across Go/HTML/CSS), a runner that hits the search API, and metrics (Recall@K, MRR)
• server/rerank/ — pluggable reranker abstraction with a tei_reranker.py implementation (calls a Text Embeddings Inference service) and a noop.py passthrough
• Docker Compose update to wire in the reranker service

---

The bi-encoder (jina-embeddings-v2-base-code) already achieves Recall@5 = 100% on this codebase, meaning the correct result is always in the top 5 without any reranking. Adding a cross-encoder on top made things worse.
The root cause: both rerankers are trained on natural language pairs, not code. They don't understand code semantics well enough to improve on a purpose-built code embedding model.

Reranking would be worth revisiting if:

• The codebase grows significantly and Recall@5 starts dropping
• You collect real usage logs and fine-tune a cross-encoder on your own (query, correct chunk) pairs
• You're willing to use an LLM (e.g. Claude) to rerank the top-5 candidates — expensive per query but genuinely code-aware

For now, the bi-encoder alone is the right choice.

NOTE: This will not be merged it just here for the reference.