parity(#219): update default embedding model to potion-code-16M-v2 #73

Summary

Upstream semble #219 (949559b) changed the default embedding model and added a score filter. csp is still on the old model.

Upstream changes

1. DEFAULT_MODEL_NAME: minishlab/potion-code-16M → minishlab/potion-code-16M-v2 (src/semble/utils.py).
2. search.py: drop chunks with a falsy combined score before rerank:

   combined_scores = {chunk: score for chunk, score in combined_scores.items() if score}

csp status

• crates/csp/src/indexing/dense.rs:25 — DEFAULT_MODEL_NAME still minishlab/potion-code-16M. Not reflected.
• crates/csp/src/search.rs:212-217 — combined scores are not filtered for 0.0 before rerank (RRF makes this near-no-op, but port for parity).

Proposed work

• Bump default model to minishlab/potion-code-16M-v2 in dense.rs.
• Ensure a model-name change invalidates the on-disk index cache (like chunk_size does in load_or_build_index) — a v1-built cache must not be reused with v2. Consider adding the model name to the manifest + _metadata_matches equivalent.
• Filter 0.0 combined scores before rerank in search.rs for behavioral parity.

Refs

• Upstream: feat: update default model to potion-code-v2 MinishLab/semble#219 (949559b)
• Sync baseline check: eacbe43..949559b, 2026-07-11