fast-jev-compaction prunes tool calls without summaries

Generative summarization is the wrong tool for agent context compaction, and deterministic pruning of tool calls is the practical alternative developers need.

• –Lossless fidelity: Preserves verbatim stack traces, error messages, and exact file paths that generative LLM summaries routinely distort or omit.
• –Low-latency scoring: Uses Jev's fast categorical judgment queries instead of generative token production, evaluating full transcripts in concurrent batches.
• –Looping vulnerability: Aggressive deletion of failed tool results risks blinding agents to previous errors, potentially causing repeated failed attempts.
• –Cache efficiency tradeoffs: Modifying middle turns in conversational history invalidates prompt prefix caching, shifting costs from token volume to cache-miss recomputations.