Name channel weight 0.5 -> 1.0; it could never reach the top k

The name channel shipped at weight 0.5 and could never fire. RRF scores a
rank-r hit at weight/(60+r); with the relevance pool at k*10 the vector channel
reaches ~1/140 at its tail, so a name-only hit at 0.5/60 lost to every vector
result down to rank ~60 and could not enter the top k. Live check after deploy:
"iSAID" still returned Icelandic census records and chip-design netlists.

Weight is now 1.0 -- equal to the vector channel, which is right for a query that
literally is a repo name. False positives are held at the trigger (identifier-
shaped tokens matched at an id boundary), not by de-weighting the channel.

The fixture test missed this because it fused a 3-item pool against 20
distractors, which passes at any weight. It now uses the real pool depth
(_rerank_pool) and a full k, and fails at 0.5.

app.py

@@ -373,7 +373,13 @@ def _bm25_ids(kind: str, text: str, n: int) -> list:
 # channel that can find them.
 _NAME_MAX_TOKENS = 3   # "iSAID instance segmentation" still carries a name token
 _NAME_MIN_CHARS = 4    # shorter tokens ("ocr", "sam") match far too much
-NAME_CHANNEL_WEIGHT = 0.5  # below the vector channel: rescue a miss, never dominate
+# Equal to the vector channel. RRF scores a rank-r hit at weight/(60+r), so the
+# vector pool (k*10) reaches ~1/140 at its tail while a name-only hit sits at
+# weight/60. At 0.5 such a hit loses to every vector result down to rank ~60 and so
+# can never enter the top k — which is the entire point of the channel. False
+# positives are held back at the trigger (identifier-shaped tokens, matched at an
+# id boundary) rather than by de-weighting the whole channel.
+NAME_CHANNEL_WEIGHT = 1.0
 
 # Hyphenated descriptors that pass the identifier test but name no repo.
 _NAME_STOPWORDS = frozenset({"real-time", "state-of-the-art", "open-source", "fine-tuned"})