[flyinprogrammer]

Closes #307

Summary

Bare identifier queries (MyService, Sinatra::Base) currently give the dense list 30% of the fused score. Setting their weight to 0 ranks them by BM25 alone; natural-language queries are unchanged.

 # semble/ranking/boosting.py
-_ALPHA_SYMBOL = 0.3  # lean BM25 for exact keyword matching
+_ALPHA_SYMBOL = 0.0  # BM25 only: exact identifier lookups gain nothing from the dense list
 _ALPHA_NL = 0.5  # balanced semantic + BM25

search(symbol query)
  fused = 0.0 × dense_rrf + 1.0 × bm25_rrf     # dense-only chunks score 0 and are dropped
  apply boosts (definition boost etc.)         # unchanged
  rerank_topk(penalise_paths = alpha < 1.0)    # 0.0 < 1.0: test/example penalties still apply

Evidence

• Test: test_resolve_alpha[MyService-None-0.0] fails on main's boosting.py and passes here. It checks the constant; the behaviour is covered by the benchmark below.
• Full suite: make test 391 passed; make lint and make typecheck clean.
• Benchmark annotations: NDCG@10 +0.12, 95% CI by repo [+0.02, +0.25]. The symbol category goes 0.959 → 0.962, and run_benchmark 0.852 → 0.854.
• Commit-message queries (Benchmark: add a commit-message query set so ranking changes aren't tuned to the annotations' style #301): no change. They're natural language, so none takes the symbol path.

Merge Danger

Door: two-way

A one-constant ranking change with no index or cache change; reverting restores the old ranking exactly.

Blast Radius: ranking

Only queries that is_symbol_query classifies as identifiers. The dense search still runs for them; its scores are just weighted 0.