evals: fix Auto Review blocking findings in step-3.7-flash and minimax-m2.7 reports

evals: add minimax-m2.7 n=20 long_context + LCB report

evals: add step-3.7-flash n=20 long_context + LCB report

evals: fix conclusion accuracy claim to match Pass 2 ground_truth results

evals: update minimax-m3 report with Pass 2 ground_truth long_context results

feat(evals): add minimax-m3 LCB + long-context benchmark results

refactor(evals): consolidate runners under headroom/evals CLI hub

refactor: update swe_bench references to model_creds in evals files

docs: update documentation to align with v0.21.x codebase

chore: add model_creds/ to .gitignore

docs: update README.md to reflect model_creds directory rename

refactor: rename swe_bench/ to model_creds/ and update all references

fix(ccr): update defaults and docs for inject_retrieval_marker

feat(ccr): disable retrieval marker by default

del deprecated evals

chore(transforms): remove twotrim files\n\n- Remove two_trim.py implementation file\n- Remove test_two_trim.py test file\n- Remove bench_two_trim.py benchmark file\n\nGenerated with [Claude Code](https://claude.ai/code)\nvia [Happy](https://happy.engineering)\n\nCo-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>\nCo-Authored-By: Happy <yesreply@happy.engineering>

feat(transforms): remove twotrim from pipeline\n\n- Remove TwoTrim transform implementation and configuration\n- Remove TwoTrim references from pipeline.py, config.py, __init__.py\n- Remove TwoTrim documentation from wiki\n- Remove test and benchmark files\n\nGenerated with [Claude Code](https://claude.ai/code)\nvia [Happy](https://happy.engineering)\n\nCo-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>\nCo-Authored-By: Happy <yesreply@happy.engineering>

feat(kompress): switch default to kompress-v2-base with weight-only int8 ONNX\n\n- Update HF_MODEL_ID and docstrings\n- Add HEADROOM_KOMPRESS_ONNX_FILENAME env var and fallback chain\n- Refactor ONNX loader to try candidates in order (int8-wo → fp32 → v1 int8)\n- Preserve existing CPU/GPU session options and logging\n\nCo-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>\nCo-Authored-By: Happy <yesreply@happy.engineering>

fix(proxy): guard _hook_ctx in post_compress when hooks enabled in cache mode

merge dev: dockerignore Rust artifacts, cost tracker, image compressor close

chore: consolidate step 3.7 eval outputs into evals/

fix(evals): send compress_user_messages=true in BeforeAfterRunner

docs(evals): record refreshed step 3.7 tier 2 results

fix(evals): send compress_user_messages=true to proxy in LiveCodeBenchRunner

feat(evals): accept explicit env_file in SuiteRunner, honor OPENAI_MODEL in LiveCodeBench

fix(evals): load repo envs and handle reasoning-only responses

fix(evals): add 429 retry with exponential backoff, extract shared helper

fix(evals): harden suite model selection and response extraction

fix(evals): strip trailing newline in LiveCodeBench stdout comparison

fix(evals): extract OpenAI text util, harden proxy logging, support stdin JSON tests

fix(evals): bump default max_tokens to 2048 for NVIDIA NIM models

fix(evals): harden LiveCodeBench runner semantics

fix(evals): harden benchmark runners and dataset loading

feat(evals): add BFCL v2, LiveCodeBench, and Long-context QA benchmarks

fix(swe-bench): default max-workers to 1

feat(swe-bench): add 25-instance benchmark subset, default to 1 worker

feat(swe-bench): wire SWE-agent prediction generation

docs(swe-bench): add benchmark guide for running with/without Headroom

feat(swe-bench): add credential smoke-tester

feat(swe-bench): per-model credential loader and runner wrapper

feat: surface TwoTrim savings in visualize and add runtime gate

fix: sort /visualize recent requests by timestamp numerically

fix: /visualize routing breakdown now reconciles to per-request totals

fix: content router detect code fences in short blocks and mixed-only fence cases

fix: /visualize history table newest-first ordering

feat: per-request compression segments breakdown in /visualize

fix: honor compress_user_messages kwarg + add pipeline diagrams to README

fix: /visualize Input Tokens Breakdown — total_input_tokens_delta IS forwarded, not original

fix: update test data for 3-element savings_history; document PrometheusMetrics divergence

fix: populate total_input_tokens in _rollup_redis_history; update docstring