Tighten local tokenizer integrity validation

[acari-git]

Problem

FileSystemLocalModelIntegrityChecker now rejects a model with no tokenizer data, but it accepts any single recognized tokenizer-related file. This can produce false-ready results for tokenizer formats that require a pair, such as vocab.json + merges.txt.

Current verification (2026-08-15)

• Missing tokenizer is correctly rejected.
• tokenizer.model, SentencePiece/Tiktoken-style files, vocab.json, and merges.txt are recognized.
• The checker does not currently validate required combinations such as BPE vocab + merges.
• LocalModelIntegrityCheckerTests pass.

Desired behavior

• Accept complete supported tokenizer layouts.
• Require paired files where the format requires them.
• Reject empty tokenizer data files.
• Keep Hugging Face IDs outside local filesystem inspection.
• Add tests for Fast Tokenizer, SentencePiece, BPE pair, Tiktoken, missing pair member, and empty tokenizer data.