Improve resumability: include model+glossary+prompt hash in cache/state key

[Congregalis]

Why

Resume/caching is great, but today the state key appears to be mostly URL + chunk. If the user changes important translation parameters (model, glossary, strictness, system prompt), a resume may silently reuse old chunks and produce a mixed output.

Proposal

• Define a run fingerprint (hash) derived from:
  • model name
  • glossary contents (or file hash)
  • strict mode settings / validation parameters
  • system prompt template version
• Store fingerprint in .transblog.state.json per URL.
• On resume: if fingerprint changed, either:
  • require -refresh (fail with actionable message), or
  • automatically treat as a new run (new state entry).

Acceptance Criteria

• No silent mixing of chunks produced under different settings
• Clear UX: either deterministic reuse or explicit retranslate
• Tests for fingerprint mismatch behavior

Notes

This is one of the highest “trust” improvements for long-running batch translations.