Performance improvements

[khezen]

Changed

• Agent runtime: dspy.ReAct → dspy.RLM in code reviewer, scope resolver, and supply chain auditor — RLM provides the same tool-using loop with improved context management
• default_max_iters raised from 3 → 5
• default_max_llm_calls raised from 5 → 8
• default_max_tokens reduced from 64000 → 32000
• summary signature defaults: max_iters=1, max_llm_calls=2
• audit signature defaults: max_iters=1, max_llm_calls=2
• Removed dead-code RLM fallback defaults (or 4 / or 8) in ContextSafe._create_rlm_fallback

Added

• Per-module max_iters and max_llm_calls overrides for Distiller and Cartographer reflection modules
  • YAML: memory.distiller.max_iters, memory.distiller.max_llm_calls (same for cartographer)
  • Env vars: MEMORY_DISTILLER_MAX_ITERS, MEMORY_DISTILLER_MAX_LLM_CALLS, MEMORY_CARTOGRAPHER_MAX_ITERS, MEMORY_CARTOGRAPHER_MAX_LLM_CALLS
  • GitHub Action inputs: memory-distiller-max-iters, memory-distiller-max-llm-calls, memory-cartographer-max-iters, memory-cartographer-max-llm-calls
  • Defaults: max_iters=1, max_llm_calls=2 (ChainOfThought modules — no tools to iterate)
• max_iters now propagated to doc reviewer, summarizer, and auditor ChainOfThought modules
• max_llm_calls now propagated to code reviewer, scope resolver, and supply chain auditor RLM agents
• GitHub Action: default-max-tokens input, summary-max-iters input, full audit-* signature inputs (audit-enabled, audit-model, audit-max-iters, audit-max-llm-calls, audit-reasoning-effort, audit-temperature), audit-memory-enabled input