erratum: G2 is consistency evidence, not a front-end check (README.md)

README.md

@@ -140,12 +140,16 @@ of any endpoint. The archived timestamps record the sequence; they cannot prove
 
 ## Independent evaluation and replication
 
-- **Second implementation (`G2`).** `independent_eval.py` reimplements decoding,
-  resampling, clip identity, the audio front end and the ranking, then runs the same
-  pinned weights. At batch 1 it agrees with the canonical results on 100% of R@1 hit/miss
-  outcomes in all eight checkpoint and direction comparisons. It still shares the
-  checkpoint's tokenizer and filterbank coefficients, since matching the weights requires
-  them; everything downstream of those is separate.
+- **Second implementation (`G2`), with an important limit.** `independent_eval.py`
+  reimplements decoding, resampling, clip identity, the audio front end and the ranking,
+  then runs the same pinned weights. At batch 1 it agrees with the canonical results on
+  100% of R@1 hit/miss outcomes in all eight comparisons. That agreement is evidence of
+  internal consistency, not an independent check of the front end: the second
+  implementation sets `is_longer=True` the same way the processor does (see the
+  correction at the top), so it reproduced the same branch routing and could not have
+  detected the defect. It also shares the checkpoint's tokenizer and filterbank
+  coefficients. Read it as "the scoring pipeline is reproducible end to end", not as
+  "the preprocessing is right".
 - **Metric parity.** `gate2.py` recomputes every shipped embedding dump two ways and finds
   0.0e+00 difference across the four canonical cells, two of which exercise multi-positive
   average precision.