Document baseline scope, clean CPU checks and retained limitations

BASELINE_JET_AMR.txt

@@ -0,0 +1,85 @@
+Jet / fixed-patch review baseline — 2026-10-15
+
+Base main: 6e3756c587ec08ab6a55c4a1a5f4671822d3f0a0
+Branch: baseline/jet-amr-20261014
+Publication target: Gitea Frank14f/CelerisLab only. This baseline does not
+advance main, publish to GitHub, or add new solver behavior.
+
+Scope
+The approved inventory contributes 43 previously tracked modifications and
+282 new files: D2 correctness/build/state repairs; planar slot/reservoir jet;
+static-circle surface actuation; one fixed 2:1 refinement patch; fine-interior
+circle integration; tests, examples and small historical validation records.
+Three reviewed repair logs are additionally retained: final-unit.log,
+final-gpu.log and simplification-checks.log. This explanation is the only new
+project document. Total: 329 changed/added paths relative to base main.
+
+Implementation: ed1a6ee54e07a5efbdc3f1880fde4cc357b10d7c (41 files)
+Tests/examples: cdbb9c42f6bfc65ec319313fb4e16cdc04dfdb95 (48 files)
+Historical evidence: 263f3b0c7addce7bacbb068f75ea707cf20b6810 (239 files)
+Historical source snapshots are evidence, not alternative runtime modules.
+
+Explicit exclusions
+The user's local changes to src/CelerisLab/lbm/kernels/config/config_grid.h,
+config_method.h, config_objects.h, config_physics.h and
+tests/postproc/experiment.ipynb are not included. Their already committed
+base-main versions remain in the tree. Their five original-user-files backups
+are excluded. Also excluded: the six pre-existing docs/design files,
+docs/multilevel_grid_architecture_analysis.md, tests/audit/body删减功能.md,
+tests/validation/multigrid/MULTIGRID_VERIFICATION_STRATEGY.md, the twenty files
+under research/audit-2026-10-09 and research/design-2026-10-09,
+research/j3-lag-2026-10-12/trajectory.h5, caches and the other twenty-two ignored
+historical output files. These exclusions preserve local originals; they are
+not deletions from the working copy. Historical manifests/reports can reference
+excluded material. In particular analyze_j3_saved_frames.py needs the omitted
+17.2 MB trajectory, and historical example images/summaries are not all bundled.
+This is a selected review baseline, not a self-contained archival data package.
+
+Fresh clean-checkout checks
+Detached checkout of 263f3b0c7addce7bacbb068f75ea707cf20b6810 was clean before
+and after both successful batches. Its four generated config headers and
+experiment notebook were byte-equal to base main; excluded untracked files
+were absent. Existing pycuda_3_10 interpreter; PYTHONPATH explicitly selected
+the clean checkout's src, which was asserted at startup. CUDA_VISIBLE_DEVICES
+was empty. No CUDA context/GPU simulation or new CFD trajectory was requested.
+
+pytest selection 1 (tests/unit): test_refinement.py, test_fine_surface.py.
+Result: 24 passed, 23 subtests passed, exit 0; pytest time 0.75 seconds.
+pytest selection 2 (tests/unit): test_build_contracts.py,
+test_context_ownership.py, test_initializers.py, test_simulation_lifecycle.py,
+test_state_regressions.py, test_surface_actuation.py, test_planar_velocity.py,
+test_slot_velocity.py, test_reservoir.py.
+Result: 152 passed, 13 subtests passed, exit 0; pytest time 5.06 seconds.
+Both used -q -p no:cacheprovider, PYTHONNOUSERSITE=1,
+PYTHONDONTWRITEBYTECODE=1, PYTEST_DISABLE_PLUGIN_AUTOLOAD=1,
+OPENBLAS_NUM_THREADS=1 and OMP_NUM_THREADS=1. Boundary-header tests use the
+existing C++ compiler; compilation-contract tests mock the CUDA compiler.
+
+The first host-contract batch was interrupted when the temporary monitoring
+script raced a removed private build directory. Its test exit is unknown;
+it is not counted as a pass. Only the monitor was corrected and that batch
+rerun; the initial geometry batch was not repeated. Project code and test
+criteria were unchanged. The final documentation-only commit does not alter
+the tested implementation, tests, configuration or historical evidence.
+
+Historical limits retained
+- J3 positive/negative full velocity steady-state criteria did not pass.
+- The local-stress implementation README records Claude's review as incomplete;
+  numerical execution is not a substitute for that implementation review.
+- repair provenance.json retains earlier 124/66 totals; final logs record
+  130 unit +13 subtests and 74 GPU tests. simplification-checks.log describes
+  an earlier phase with known XFAILs, not a new result on this baseline.
+- The 24-coarse-step joint run demonstrates startup disturbances crossing the
+  interface, potentially acoustic. It does not establish a developed jet,
+  physical accuracy, convergence order or long-term stability. Raw FP32 mass
+  residuals remain nonzero despite the independently measured rounding ledger.
+- Fixed-patch support is not dynamic/block-level AMR, moving rigid bodies,
+  complete 3D or multi-GPU support. Historical full suites were not rerun here.
+
+Publication screening
+Approved candidate bytes and the three logs were checked for common embedded
+credential/private-key patterns and unrelated private-location indicators;
+unchanged tracked blobs and staged content were also checked. No findings were
+detected. This bounded check is not a general security audit. Excluded files
+were hash-protected. No credentials, accounts, environments or remotes were
+modified by the baseline preparation.