Calibrate spam prompt confidence

[dtump]

Summary

Make a small, generic calibration improvement to the shared spam-classification prompt without changing Inspamity's scoring contract.

Changes

• Explicitly define confidence as confidence in the selected yes/no verdict; for example, is_spam=no with confidence 90 means strong confidence that the message is not spam and therefore maps to a large negative Rspamd score.
• Prevent unsupported claims about recipient subscription history, domain reputation, or unfamiliar domains.
• Treat absent preprocessed SPF/DKIM information as unknown rather than failed.
• Clarify that authentication problems, promotional language, urgency, tracking infrastructure, image count, and awkward language are weak signals when isolated.
• Require identity or URL mismatches to be considered with other suspicious evidence.
• Add tests preserving these prompt and confidence semantics.

The prompt remains concise and generic: it contains no local domains, regional brands, sender allowlists, or user-specific preferences.

Validation

• .venv/bin/ruff check .
• .venv/bin/ruff format --check .
• .venv/bin/pytest — 100 passed, 1 skipped

No live provider calls were made. Model-specific behavior should be compared before production deployment because prompt changes can still cause classification regressions.