feat(scripts): build a draft vocabulary mapping without regenerating features

[fg11991]

Motivation

Choosing a draft vocabulary currently requires a full hidden-state capture first: process_token_dict_to_mappings runs inside prepare_hidden_states.py. Trying a second draft_vocab_size, or building a mapping for a corpus whose features were captured before a mapping was needed, means paying for the capture again — hours of GPU time on a 600k-sample corpus to answer a question about token frequencies.

What's in it

scripts/build_vocab_mapping.py derives the mapping from either source, with no GPU involved:

• --data-path counts loss-bearing tokens from the source JSONL, applying the same chat template and truncation the trainer would;
• --hidden-states-path counts them from already prepared features.

Both stream: the tally is accumulated per shard instead of materializing a token column for the whole corpus, and per-corpus counts are cached under a key that includes the corpus identity, so sweeping several draft_vocab_size values re-reads nothing.

Two supporting changes:

• specforge/training/vocab_mapping.py → specforge/data/. Scripts may not import specforge.training (tests/test_runtime/test_package_architecture.py enforces this), and counting tokens in prepared features is a data concern that the trainer merely happens to be the first caller of. The module itself is unchanged.
• process_token_dict_to_mappings builds t2d by scattering into a zeroed mask instead of testing membership in a Python list per target id. The old form is quadratic in the vocabulary: tens of seconds at V=248320, K=64000, inside a phase that prints nothing, for a result that takes milliseconds.

Testing

tests/test_scripts/test_build_vocab_mapping.py covers both input paths, the cache-identity key, gzipped features, and — the part that actually matters — that the written file loads into a real EAGLE3 draft model's buffers and keeps the frequent tokens. tests/test_data/test_vocab_mapping_construction.py covers the scatter rewrite against the previous list-based result.

Full suite run on CPU (no CUDA available on the dev box): no test fails that does not already fail on main at the same commit.