fix: preserve truncated reward-memory evidence semantics

[LIHUA919]

This is a focused follow-up for huangruiteng#3829 at exact head e278d1c.

Summary:

• add a sparse evidence-label joint summary so truncated readback validation can rederive the strongest tier, effective label, confidence, and review state;
• replace exception-message substring classification with typed Stage-1 validation reason propagation;
• add the 257-observation harmful-to-helpful tamper regression plus exact reason-code coverage;
• update the English and Chinese projection contracts.

Validation:

• 142 Reward Memory tests passed;
• Reward Memory architecture and utility-reducer smokes passed;
• Ruff lint and format checks passed;
• targeted mypy passed;
• LoopX premerge standard gate passed 17/17 with no manual holds;
• public/private boundary scan passed.

The change remains read-only and does not add ranking, provider writeback, memory lifecycle, quota, scheduler, or action authority.