v0.1.4

New Features

• Compression ratio guard (#20) — detects when compression loses too much content and records compression_strategy / ratio_violation in metrics
• Compression feedback tool (#21) — stm_compression_feedback lets agents report information loss as a learning signal; stm_compression_stats shows aggregated feedback
• Ratio guard fallback (#22) — boundary-aware truncate fallback when compression overshoots the retention floor
• 3-tier fallback ladder — progressive → hybrid → truncate. New hybrid tier preserves document structure (head + TOC) for content with ≥ 3 headings that is too small for progressive chunking
• Per-tool retention_floor — override the global dynamic retention scaling per server or per tool in stm_proxy.json
• Compression auto-tuner — stm_tuning_recommendations MCP tool analyses proxy metrics and produces per-tool recommendations (budget increase/decrease, strategy pinning, feedback-driven strategy changes)
• Nested Langfuse sub-spans — proxy_call_clean, proxy_call_compress, proxy_call_surface, proxy_call_index, proxy_call_cache_hit nested under the top-level proxy_call observation
• Langfuse sampling — MEMTOMEM_STM_LANGFUSE__SAMPLING_RATE (0.0–1.0) to reduce tracing volume for high-throughput deployments

Improvements

• LLM fallback signal recorded in compression_strategy metric (e.g. llm_summary→privacy_fallback)
• Embedding scorer fallback count exposed in CallMetrics.scorer_fallback
• SKELETON body_trimmed_chars metadata in compression footer
• Convention suffix in proxied tool descriptions for progressive/selective delivery
• Startup warning when compression != none but auto_index is disabled

Docs

• 3-tier fallback ladder diagram in compression.md
• retention_floor config reference in configuration.md
• 7 → 10 MCP tools in cli.md (stm_compression_feedback, stm_compression_stats, stm_tuning_recommendations)
• Langfuse nested span table and sampling config in operations.md

Testing

• 963 automated tests (up from 800)

Full Changelog: v0.1.3...v0.1.4