Same recipe as a2B -- seed_altfix2, 49,152 pieces, evict all, compound floor 4,000, twins 2,000/2,000. Only the fit corpus changed: 700 Mchar drawn proportional to the training budget from the re-cleaned, near-duplicate-filtered corpus.
Paradigm sharing rises. One stem row + suffixes 44.1% -> 45.4%, prefix kept 72.4% -> 73.1%, lemma is one token 54.3% -> 55.1%, at an unchanged 3.04 tokens/form.
benchmark.py: bigcode/starcoderdata moved 'python' from a config to a data directory upstream, so the script raised BuilderConfig not found on its last corpus. One-line fix.
