refactor: Unified prompt composition, structured log analysis, parallel validator tracing, and intent hydration fixes

[sortedcord]

This pull request compiles all architectural improvements, refactoring, and fixes implemented in this branch to unify prompt generation, structure logs, add parallel validator checks, and fix entity/dialogue hydration.

1. Unified Prompt Composition and Generic Interfaces

• Centralized LLM prompt types by relocating PromptComponent and PromptBreakdown from @omnia/voice to @omnia/llm (packages/llm/src/llm.ts).

• Standardized prompt construction by implementing the generic IPromptBuilder<TArgs> interface across all prompt builders:

  • ActorPromptBuilder (packages/actor/src/actor-prompt-builder.ts) implements IPromptBuilder<[WorldState, Entity]>.
  • IntentDecoderPromptBuilder (packages/intent/src/intent-prompt-builder.ts) implements IPromptBuilder<[WorldState, string, string, Intent[]]>.
  • HandoffPromptBuilder (packages/memory/src/handoff-prompt-builder.ts) implements IPromptBuilder<[Entity, BufferEntry[], Date]>.
  • LLMValidatorPromptBuilder (packages/architect/src/llm-validator-prompt-builder.ts) implements IPromptBuilder<[WorldState, Intent]>.

• Decoupled prompt composition from LLM execution, providing a consistent and reusable prompt-building architecture across the codebase.

2. Structured Simulation Logging and Reusable Prompt Visualization

• Replaced fragile regex-based parsing of provider lastCall output in turn-executor.ts and alias-handoff.ts.

• Introduced structured prompt logging by storing prompt breakdowns as token-labeled JSON component arrays.

• Added a reusable PromptAnalyzer component (apps/gui/src/components/play/PromptAnalyzer.tsx) that:

  • Renders color-coded prompt sections.
  • Displays System, World, Memory Context, and Candidate Buffers independently.
  • Preserves backward compatibility through legacy string-splitting fallbacks.

• Refactored both PromptModal and HandoffModal to delegate prompt rendering and visualization responsibilities to PromptAnalyzer, significantly simplifying modal logic.

3. Parallel Intent Validation Logging and Visualization

• Separated raw intent decoding from the validation pipeline.

• Preserved decoded intents in LogEntry.decodedIntents, allowing the interface to display model output before validation occurs.

• Redesigned PromptSwitcher to visually represent the parallel validation architecture:

  • Primary execution flow remains horizontal.
  • Individual validator executions are stacked vertically beneath the main pipeline.

• Added dynamic validation tabs for each LLM validation execution.

• Each validation view displays:

  • Prompt token breakdown
  • Model response
  • Validation status

• Added a dedicated "Bypassed LLM Validation" view to clearly explain scenarios where validation is intentionally skipped, such as dialogue and thought intents.

4. Grammar Preprocessing and Dialogue Preservation

• Introduced the expandContractions preprocessor (packages/voice/src/contractions.ts).
• Expands narrator contractions before intent decoding while intentionally ignoring quoted dialogue.
• Prevents accidental modification of spoken dialogue while improving parser consistency for narration.

5. Hydration Alias Fallback Improvements

• Updated IntentTag in InteractView.tsx to receive simulation entities during rendering.

• Improved alias hydration by:

  • Pre-populating alias mappings with every entity's default name.
  • Overlaying player-specific aliases afterward.

• Ensures unmapped entities resolve to their canonical names (for example, "Bob" or "Bill") instead of the generic placeholder "an unfamiliar figure."