HEG BRep Component Identification — Distribution Bundle
=========================================================

This folder is a self-contained Windows distribution. It runs without any
pre-installed Python or conda. Drop the whole folder next to your viewer's
installer; the viewer launches `heg_brep_service.bat` as a child process and
talks to it over localhost HTTP.


Contents
--------
heg_brep_service.bat   Launches the classification service. Prints
                         `READY port=<N>` to stdout once models are loaded.
                         Use `--device cpu` (default) or `--device cuda`.
heg_brep_batch.bat     Optional batch CLI: STEP folder -> Excel report.
                         Useful for offline regression testing.
python\                 Conda-packed Python 3.10 environment (~2 GB).
                         Includes pythonocc-core, occwl, torch (CPU),
                         torch_geometric, fastapi, uvicorn, numpy, MKL.
heg_brep\               The classification package (model, inference,
                         extraction, server, batch CLI).
BRepExtractor\          STEP -> NPZ feature extraction pipeline.
models\                 Three checkpoints used by the two-pass classifier:
                         pass1.pt (parent), elbow.pt, tee.pt.
csharp_sample\          Reference C# integration: HegBrepClient.cs is a
                         drop-in async client; Program.cs is a smoke test
                         (build with `dotnet build`).


Quick smoke test
----------------
From this folder, in a regular cmd.exe:

  heg_brep_batch.bat C:\path\to\step_folder C:\path\to\out.xlsx

Or run the service standalone and poke it with curl:

  heg_brep_service.bat
    (waits ~10-15s; prints `READY port=51571`)
  curl http://127.0.0.1:51571/health
  curl -X POST http://127.0.0.1:51571/classify ^
       -H "Content-Type: application/json" ^
       -d "{\"step_path\":\"C:\\path\\to\\part.step\"}"


Service HTTP API
----------------
GET  /health
       -> {"status":"ok","models_loaded":true,"device":"cpu","uptime_sec":...}

POST /classify           Single STEP file -> classification result.
       Body:  {"step_path": "C:\\full\\path\\to\\part.step",
               "npz_keep_dir": "optional\\persist\\dir"}
       Resp:  {"status":"ok",
               "final_label": "4_tee_wf",
               "final_conf": 0.9997,
               "route": "tee",
               "pass1_argmax": "tee", "pass1_conf": 1.0,
               "pass2_argmax": "4_tee_wf", "pass2_predicted": "4_tee_wf",
               "pass2_conf": 0.9997,
               "npz_path": "...", "step_path": "..."}

POST /classify_batch     Multiple STEPs in one call. Same JSON schema,
                         but body is `{"step_paths": ["...", "..."]}`.

POST /shutdown           Graceful exit. The host (your viewer) should call
                         this on exit so the python.exe child doesn't linger.


Performance budget (RTX 3050 laptop, --device cpu)
--------------------------------------------------
Service startup:       ~10-15s   (OCC + torch + 3 models cold-load)
First /classify call:  ~1.0s     (single STEP extract + inference)
Warm /classify calls:  ~0.8s     (warm OCC stack, varies with STEP size)

Switch to `--device cuda` for ~5-10x faster inference if the target machine
has CUDA 12.x drivers + a supported GPU. With CPU torch (default), no NVIDIA
runtime is required.


Output label space
------------------
Pass-1 (parent classifier): elbow | tee | pipe | miscellaneous
Pass-2 elbow specialist:    1_elbow_wf | 2_elbow_pef | 3_elbow_sf | 8_elbow_misc
Pass-2 tee specialist:      4_tee_wf | 5_tee_pef | 6_tee_sf | 9_tee_misc

For pipe / miscellaneous routes the response has `final_label="random"` with
the pass-1 confidence (no specialist exists for those families).


C# integration
--------------
See csharp_sample/HegBrepClient.cs for a drop-in async client. Typical usage:

    var client = new HegBrepClient();
    await client.StartAsync(@"C:\Program Files\YourViewer\heg_brep\heg_brep_service.bat");
    // Later, when user clicks "Identify":
    var r = await client.ClassifyAsync(@"C:\tmp\selection.step");
    label.Text = $"{r.FinalLabel} ({r.FinalConf:P1})";
    // On viewer exit:
    await client.StopAsync();

The client spawns the .bat as a child process, reads the `READY port=<N>`
line from its stdout, then POSTs to http://127.0.0.1:<N>/classify. Stderr is
forwarded to Debug.WriteLine for your logger to capture.


Troubleshooting
---------------
* "DLL load failed" on first run.
    The bat sets PATH to include python\Library\bin where MKL / OCC DLLs
    live. If you launch python.exe directly without the bat, you'll hit this.

* Cold service start takes 30+ seconds.
    The first run after extracting the zip is slow because Windows loads each
    DLL from disk for the first time. Subsequent starts are ~5-10s once the
    OS file cache is warm.

* "Bodies which are not closed are not supported".
    Upstream extractor limitation. Some STEP files (typically Inventor
    exports without explicit BREP closure) cannot be processed. The HTTP
    response is {"status": "extraction_failed", "error": "..."}.

* `KMP_DUPLICATE_LIB_OK=TRUE` warning.
    Set automatically. Coexists fine with both OpenMP runtimes; the warning
    is overly conservative for inference.