Preserve backbone precision and extend smoke coverage

[AngeLouCN]

Ordinary decision scoring retains unused generation caches and casts the entire sequence of hidden states to FP32, although the pointer head reads only selected vectors. Route text, media and prefix execution through BackboneAdapter.forward, retain the backbone's native hidden-state dtype, and cast only the vectors needed by the pointer head. Ordinary scoring defaults to use_cache=False; explicit prefix reuse continues to request a cache.

Extend the backbone smoke CLI with separate frozen-weight precision and decision-mode options. Handle tokenizers that reuse existing delimiter tokens without requiring jevany_token_schema metadata, reject incompatible direct-token training flags, and use a tiny GLM fixture with dimensions previously verified for BF16 SDPA on A100.

Validation on the restored two-commit branch:

• CPU regression: python -m pytest tests -m 'not server' -q -ra --strict-config --strict-markers — 432 passed, 37 skipped, 7 deselected.
• python -m build --no-isolation — source distribution and wheel both built successfully.
• Optional CUDA checks are skipped in the CPU suite; no new GPU run was performed for this revision.

Base: main at 625aed9.

Candidate-token latency optimization and its accuracy/latency results are reviewed separately in #9.