CLAUDE: Stage 2 — JNA convert_direct binding + Stage2 validation driver

TpJna: add tp_convert_direct_selftest binding. Stage2: invoke it against the
tile_processor_gpu/clt golden, report num_active_tiles + max|CLT-golden|.
PASS on 5060 Ti: 5120 active tiles, relative error ~8.85e-6 vs golden (first
real kernel execution + CDP via the native shim, no JCuda).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

src/main/java/com/elphel/imagej/gpu/jna/Stage2.java

+package com.elphel.imagej.gpu.jna;
+
+import com.sun.jna.Native;
+import com.sun.jna.Pointer;
+
+/**
+ * Stage-2 validation: the convert_direct milestone. Drive the full direct-CLT path natively
+ * (NVRTC module + driver-API launches, no JCuda) through the JNA boundary, and compare the
+ * computed CLT to the tile_processor_gpu golden (clt/aux_chnN.clt). This is the first real
+ * kernel-execution test and the first CDP (calculate_tiles_offsets) run on Blackwell via the shim.
+ *
+ * Run:
+ *   java -Djna.library.path=<dir with libtileproc.so> -cp target/classes:<jna.jar> \
+ *        com.elphel.imagej.gpu.jna.Stage2 [data_root] [lwir]
+ *   data_root must contain clt/ (default /home/elphel/git/tile_processor_gpu); lwir default 1.
+ * By Claude on 2026-06-25.
+ */
+public class Stage2 {
+    public static void main(String[] args) {
+        String root = (args.length > 0) ? args[0] : "/home/elphel/git/tile_processor_gpu";
+        int    lwir = (args.length > 1) ? Integer.parseInt(args[1]) : 1;
+        String srcdir = "/home/elphel/git/tile_processor_gpu/src";
+        String devrt  = "/usr/local/cuda/targets/x86_64-linux/lib/libcudadevrt.a";
+
+        System.out.printf("Stage2 (convert_direct via JNA): root=%s lwir=%d%n", root, lwir);
+        TpJna lib = Native.load("tileproc", TpJna.class);
+        Pointer m = lib.tp_create_module(srcdir, devrt);
+        if (m == null) { System.out.println("FAIL: tp_create_module:\n" + lib.tp_last_error()); System.exit(1); }
+
+        double[] maxErr = new double[1];
+        int[]    nActive = new int[1];
+        int rc = lib.tp_convert_direct_selftest(m, lwir, root, maxErr, nActive);
+        if (rc != 0) { System.out.printf("FAIL: selftest rc=%d: %s%n", rc, lib.tp_last_error()); System.exit(2); }
+
+        // CDP can reorder FP accumulation, and NVRTC vs the harness's offline nvcc may differ slightly,
+        // so this is a closeness check, not bitwise. Report the number; judge from it.
+        boolean ok = (maxErr[0] >= 0) && !Double.isNaN(maxErr[0]) && (maxErr[0] < 1.0) && (nActive[0] > 0);
+        System.out.printf("  convert_direct: num_active_tiles=%d  max|CLT-golden|=%.6g  %s%n",
+                nActive[0], maxErr[0], ok ? "PASS" : "REVIEW");
+        lib.tp_destroy_module(m);
+        System.out.println(ok ? "RESULT: PASS (convert_direct matches golden via JNA)" : "RESULT: REVIEW (see maxErr)");
+        System.exit(ok ? 0 : 3);
+    }
+}

src/main/java/com/elphel/imagej/gpu/jna/TpJna.java

@@ -36,4 +36,12 @@ public interface TpJna extends Library {
     int     tp_get_rot_deriv(Pointer instance, float[] out);
     /** Free the instance device buffers. */
     void    tp_destroy_instance(Pointer instance);
+
+    // ---- Stage 2: convert_direct selftest ----
+    /** Run the full convert_direct path (mirrors TpHostGpu allTests convert chain) against the
+     *  tile_processor_gpu/clt golden data under data_root. lwir!=0 -> 16-cam LWIR set.
+     *  Writes max |CLT-golden| into outMaxErr[0] and the kernel-computed active-tile count into
+     *  outNumActive[0]. Returns 0 on success. */
+    int     tp_convert_direct_selftest(Pointer module, int lwir, String data_root,
+                                       double[] outMaxErr, int[] outNumActive);
 }