Optimize benchmark execution and reset evaluation baseline

[brunovicco]

Summary

• separate the canonical Gemini embedding benchmark from the local Qwen variant
• add a deterministic stratified validation/test split
• persist embeddings and safely reuse completed dense, sparse, and GraphRAG indexes
• make strategy selection and benchmark resume behavior effective
• prevent concurrent runs from mutating shared indexes
• add stage timing, token usage, configurable cost estimates, and metric breakdowns
• clarify the relevant architecture decisions and benchmark documentation
• remove previous results and invalidate the old embedding evaluation baseline

Motivation

Full benchmark runs were repeatedly rebuilding embeddings and indexes, making indexing unnecessarily slow and costly. Previous results also predated the corrected split and execution behavior and could no longer serve as a valid comparison baseline.

Redis was not introduced because it would not directly replace the current pgvector and OpenSearch responsibilities. Persistent embedding caching and safe index reuse provide the immediate performance improvement without changing the storage architecture.

Validation

• full quality gate passed
• 480 tests passed, 20 deselected
• 84.26% test coverage
• Ruff, formatting, Mypy, architecture, governance, and security checks passed
• dependency audit reported no known vulnerabilities outside configured ignores
• git diff --check passed

Operational notes

• make bench-live uses the canonical Gemini embedding configuration
• make bench-live-local uses local Qwen embeddings
• local NVIDIA hosts can enable CUDA by setting embedding.device: cuda
• live benchmark execution was not performed because it would incur external provider costs