Split: engine only here; the environment moves to BlameGraph

[AdvayMonga]

This repo becomes the engine and its stack. Everything that runs, measures or judges the engine moves to BlameGraph: the lab, the corpus, the knowledge base and the GPU VMs. Pair PR: AdvayMonga/BlameGraph#3. The two should merge together.

Removed here (moved to BlameGraph)

• lab/: session loop, agent, tools, ledger, workspace, budget, sandbox/grader, canaries, profiler, vm.py with the Verda/Nebius/Crusoe providers, vm-setup.sh
• corpus/ and scripts/corpus/, knowledge/, ENVIRONMENT.md, tests/test_lab_*.py, tests/lab_fixtures.py
• scripts/bench/roofline.py (going into BlameGraph validity/ to be generalized)

Removed, superseded by BlameGraph's load regimes

bench_serving.py, load_test.py, prompt_bank.py, plot_load_test.py, baseline_benchmark.py, the eviction/fairness/KV-pressure benchmarks, scripts/probes/, scripts/tools/test_batch_cache.py and test_scheduler.py, and their tests (test_bench_serving, test_load_test, test_load_workloads).

Kept

src/ (engine, control plane), engine tests, monitoring/, docs/papers/, scripts/bench/tune_triton_launch.py (now writes timing evidence to runs/timing/), scripts/gpu_tests/checks.py, scripts/tools/smoke_custom.py, scripts/hooks/pre-push.

Other changes

• CI keeps the engine lane only; the corpus provenance lane moves to BlameGraph's CI.
• pyproject.toml drops the agents and corpus extras; uv.lock is regenerated accordingly.
• tests/test_qwen3_moe_parity.py::test_chat_template_disables_thinking no longer imports the lab. It checks the template with enable_thinking=False directly; the fingerprint half is covered by BlameGraph's test_lab_chat_template.py.
• README, CLAUDE.md, CONTRIBUTING.md, scripts/README.md and the PR template now point at BlameGraph for anything environment.
• Tag archive/pre-split (= ef9ce63) holds the combined tree.

Verified

• ruff check . clean. Engine suite: 360 passed, 37 skipped.
• From BlameGraph against this branch: the lab's test tool (lint + 360 tests on the agent's pristine copy) and profile tool (bundle written) both work, running this repo's interpreter.

Notes

• Open PRs corpus: three mixed classes crossing regimes, per-request regime labels #97 (corpus mixed classes) and knowledge: Qwen3-30B-A3B quantized checkpoints on vLLM 0.30; H200 venue image differences #99 (knowledge entries) were carried into BlameGraph as-is. They can be closed here once this merges.
• Local state to move by hand: lab/ledger/ and lab env vars / credentials.

🤖 Generated with Claude Code