feat(ai): adaptive embedding batch sizing circuit breaker and timeout fallback

[dan-petty]

Problem Statement

Distributed OpenTelemetry traces indicate ai.rag.ollama_embed_batch repeatedly encounters 15.0s ReadTimeout: timed out exceptions (75s+ total latency) during dense repository indexing, stalling RAG operations on overloaded local Ollama nodes.

Proposed Changes

• Dynamic batch partitioning in devops_cli.ai.rag.embeddings adapting batch size dynamically (halving from 32 down to 16, 8, 4 on latency degradation or timeout).
• Exponential backoff with jitter on transient network timeouts.
• Pre-flight Valkey SHA-256 chunk cache checks before remote embedding dispatch.
• Automatic fallback to single-chunk requests when batches fail.

Acceptance Criteria

• Zero uncaught ReadTimeout exceptions during full workspace re-indexing.
• Average batch latency reduced below 2.0s.
• 100% unit test coverage of batch adaptation logic in tests/test_rag_embeddings.py.