MAINT Decompose scorer execution policy and result finalization

[Haseeb Nazir (iamhaseebn)]

Description

MessageScorer currently handles message substitution, skipping, objective inference, execution errors, fallbacks, and evidence anchoring in one flow. This separates that work into preparation, execution, and finalization steps while keeping the policy in the scorer layer.

The branch has also been brought onto current main. The conflict resolution preserves the newer scorer architecture: capability-based role handling and undetermined results remain in MessageScorer, while shared score validation and persistence remain in Scorer and occur exactly once.

The refactor preserves behavior for structured refusals, blocked partial content, role and error filtering, explicit and inferred objectives, scorer-owned blocked responses, exception propagation, fallback scores, ephemeral links, persisted evidence, and downstream callers.

Closes #2413.

Tests and Documentation

• Message scorer tests: 67 passed
• Scorer and downstream caller coverage: 339 passed
• Complete scorer suite: 1,760 passed
• Complete unit suite: 17,176 passed, 10 skipped
• Feature diff coverage: 100%
• All-files pre-commit checks passed

No documentation changes are needed because this is an internal refactor with no public API or intended behavior changes.

[Haseeb Nazir (iamhaseebn)]

@microsoft-github-policy-service agree

[Richard Lundeen (richlundeen)]

Haseeb Nazir (@iamhaseebn) this is a good change, but I have a similarly stacked PR to progress with this. They do similar things and right now the conflicts aren't worth it. I'd wait at least a month before looking into non-bug scoring refactor changes.

Thanks for taking a look!