The gateway had a SEPARATE compression system ('session hygiene')
with hardcoded thresholds (100k tokens / 200 messages) that were
completely disconnected from the model's context length and the
user's compression config in config.yaml. This caused premature
auto-compression on Telegram/Discord — triggering at ~60k tokens
(from the 200-message threshold) or inconsistent token counts.

Changes:
- Gateway hygiene now reads model name from config.yaml and uses
  get_model_context_length() to derive the actual context limit
- Compression threshold comes from compression.threshold in
  config.yaml (default 0.85), same as the agent's ContextCompressor
- Removed the message-count-based trigger (was redundant and caused
  false positives in tool-heavy sessions)
- Removed the undocumented session_hygiene config section — the
  standard compression.* config now controls everything
- Env var overrides (CONTEXT_COMPRESSION_THRESHOLD,
  CONTEXT_COMPRESSION_ENABLED) are respected
- Warn threshold is now 95% of model context (was hardcoded 200k)
- Updated tests to verify model-aware thresholds, scaling across
  models, and that message count alone no longer triggers compression

For claude-opus-4.6 (200k context) at 85% threshold: gateway
hygiene now triggers at 170k tokens instead of the old 100k.

Major updates to reflect the current OBLITERATUS codebase:

- Change default recommendation from 'informed' (experimental) to
  'advanced' (reliable, well-tested multi-direction SVD)
- Add new CLI commands: tourney, recommend, strategies, report,
  aggregate, abliterate (alias)
- Add --direction-method flag (diff_means, svd, leace)
- Add strategies module (embedding/FFN ablation, head pruning,
  layer removal)
- Add evaluation module with LM Eval Harness integration
- Expand analysis modules from 15 to 28
- Add Apple Silicon (MLX) support
- Add study presets (quick, jailbreak, knowledge, etc.)
- Add --contribute, --verify-sample-size, --preset flags
- Add complete CLI command reference table
- Fix torch property name: total_mem -> total_memory (caught
  during live testing)

Tested: Successfully abliterated Qwen2.5-0.5B-Instruct using
'advanced' method — refusal rate 0.4%, coherence 1.0, model
responds without refusal to test prompts.

Added pitfalls discovered during live abliteration testing:
- Models < 1B have fragmented refusal, respond poorly (0.5B: 60%→20%)
- Models 3B+ work much better (3B: 75%→0% with advanced defaults)
- aggressive method can backfire on small models (made it worse)
- Spectral certification RED is common even when refusal rate is 0%
- Fixed torch property: total_mem → total_memory