chore(applet): strip diagnostic INSes from prod CAP (M2F.3)

Flip DIAGNOSTICS_ENABLED static-final boolean from true to false so
the production j3r452 CAP rejects INS_DIAG_HMAC/ECDH/ECDSA_SIGN/GCM
(0xD0-0xD3) with SW_INS_NOT_SUPPORTED. javac's constant-folding
elides the diagnostic dispatch branch in process(); the JC converter
keeps the unreachable processDiag method body and DIAG_KEY_32 etc.
arrays so the byte saving is modest (the JC converter does not do
whole-program DCE on private fields/methods).

CAP size (j3r452): before 101008 B -> after 100989 B (delta -19 B,
0.02% reduction). This is smaller than the original ~2-3 KB estimate
in the plan -- the estimate assumed javac/JC-converter DCE would
prune the diag arrays and methods, but neither does. Reclaiming the
~2 KB diag-vector + diag-method footprint requires extracting the
diag code to a separate class and excluding it from the production
profile, which is out of scope for M2 (file-level surgery, not a
constant flip). The runtime SW_INS_NOT_SUPPORTED rejection still
delivers the security objective: no diag opcodes execute on a prod
build.

For CI/bench builds that need the diag opcodes, the constant can be
locally flipped to true (no test changes required; no diag tests in
the suite reference INS_DIAG_*).

Regression: 147 tests, 0 failures, 3 pre-existing GCM errors (no
new failures).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

applet/src/main/java/com/dangerousthings/aliro/AliroApplet.java

@@ -46,7 +46,7 @@ public class AliroApplet extends Applet {
     // AUTH session scratch). The diagnostic keypair is allocated separately
     // from the protocol's ephemeral keypair, so calling diag mid-transaction
     // cannot clobber an in-flight AUTH0/AUTH1.
-    private static final boolean DIAGNOSTICS_ENABLED = true;
+    private static final boolean DIAGNOSTICS_ENABLED = false;
 
     private static final byte INS_DIAG_HMAC       = (byte) 0xD0;
     private static final byte INS_DIAG_ECDH       = (byte) 0xD1;