CLAUDE: Stage 3 — JNA imclt_rbg_all binding + Stage3 validation driver

TpJna: add tp_imclt_selftest binding. Stage3: run convert_direct + imclt_rbg_all,
report CLT and RBG max error. PASS on 5060 Ti: RBG relative ~1.31e-5 vs golden.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

src/main/java/com/elphel/imagej/gpu/jna/Stage3.java

+package com.elphel.imagej.gpu.jna;
+
+import com.sun.jna.Native;
+import com.sun.jna.Pointer;
+
+/**
+ * Stage-3 validation: imclt_rbg_all. Run convert_direct then the inverse-CLT RBG reconstruction
+ * natively (NVRTC + driver-API, no JCuda) through JNA, and compare the de-pitched RBG output to
+ * the tile_processor_gpu golden (clt/aux_chnN.rbg). Builds directly on the Stage-2 setup.
+ *
+ * Run:
+ *   java -Djna.library.path=<dir with libtileproc.so> -cp target/classes:<jna.jar> \
+ *        com.elphel.imagej.gpu.jna.Stage3 [data_root] [lwir]
+ * By Claude on 2026-06-25.
+ */
+public class Stage3 {
+    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("Stage3 (convert_direct + imclt_rbg_all 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[] cltErr = new double[1], rbgErr = new double[1];
+        int[]    nActive = new int[1];
+        int rc = lib.tp_imclt_selftest(m, lwir, root, cltErr, rbgErr, nActive);
+        if (rc != 0) { System.out.printf("FAIL: selftest rc=%d: %s%n", rc, lib.tp_last_error()); System.exit(2); }
+
+        // Closeness check (NVRTC vs offline-nvcc + CDP reorder), not bitwise.
+        boolean ok = (rbgErr[0] >= 0) && !Double.isNaN(rbgErr[0]) && (rbgErr[0] < 1.0) && (nActive[0] > 0);
+        System.out.printf("  convert_direct: num_active=%d  max|CLT-golden|=%.6g%n", nActive[0], cltErr[0]);
+        System.out.printf("  imclt_rbg_all:  max|RBG-golden|=%.6g  %s%n", rbgErr[0], ok ? "PASS" : "REVIEW");
+        lib.tp_destroy_module(m);
+        System.out.println(ok ? "RESULT: PASS (imclt_rbg_all matches golden via JNA)" : "RESULT: REVIEW (see maxErr)");
+        System.exit(ok ? 0 : 3);
+    }
+}

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

@@ -44,4 +44,10 @@ public interface TpJna extends Library {
      *  outNumActive[0]. Returns 0 on success. */
     int     tp_convert_direct_selftest(Pointer module, int lwir, String data_root,
                                        double[] outMaxErr, int[] outNumActive);
+
+    // ---- Stage 3: convert_direct + imclt_rbg_all ----
+    /** Run convert_direct then imclt_rbg_all; compare CLT to clt/aux_chnN.clt (outCltErr[0]) and
+     *  the de-pitched RBG to clt/aux_chnN.rbg (outRbgErr[0]). Returns 0 on success. */
+    int     tp_imclt_selftest(Pointer module, int lwir, String data_root,
+                              double[] outCltErr, double[] outRbgErr, int[] outNumActive);
 }