fix(mcu): PR-AB.a — move vector_0 out of inner beam_pos loop

runRadarPulseSequence used to fire vector_0 (broadside reference)
between every matrix1 and matrix2 pattern, i.e. 15 times per azimuth.
That dwell × 8 ms × 15 = 120 ms per azimuth × 50 azimuths = ~6 s of
the 18.4 s revisit time was burned on redundant broadside frames.

Pull vector_0 out of the loop and fire it once per azimuth before the
sweep. Each azimuth now produces 1 broadside frame + 30 steered frames
(matrix1 + matrix2 across 15 beam_pos), down from 15 + 30 = 45 frames.
Revisit time drops from 18.4 s to ~12.8 s (31% improvement).

If multiple per-position broadside frames are ever needed, gate them
behind a runtime switch — the comment block flags this.

test_bug16_runradar_shadows_globals updated to mirror the new
1-outside + 2-inside m-counter pattern; 13/13 PASS, full MCU
regression 51/0 + 34/0.

9_Firmware/9_1_Microcontroller/tests/test_bug16_runradar_shadows_globals.c

@@ -59,13 +59,20 @@ static void run_buggy(void)
     (void)m; (void)n; (void)y;
 }
 
-/* Post-fix: same body, no local redeclaration — references globals. */
+/* Post-fix: same body, no local redeclaration — references globals.
+ * PR-AB.a moved vector_0 out of the inner loop, so the m advance is now
+ *   1 (vector_0, before loop) + 2 × 15 (matrix1+matrix2 in loop) = 31 increments
+ * per azimuth, instead of the prior 3 × 15 = 45. The wrap behavior
+ * (g_m wraps on every iteration's matrix2 add) is unchanged. */
 static void run_fixed(void)
 {
+    /* PR-AB.a: vector_0 broadside reference, 1× per azimuth (was inside loop). */
+    g_m += m_max / 2;
+    if (g_m > m_max) g_m = 1;
+
     for (int beam_pos = 0; beam_pos < 15; beam_pos++) {
-        g_m += m_max / 2;
-        g_m += m_max / 2;
-        g_m += m_max / 2;
+        g_m += m_max / 2;  /* matrix1 (negative-θ scan) */
+        g_m += m_max / 2;  /* matrix2 (positive-θ scan) */
         if (g_m > m_max) g_m = 1;
 
         g_n++;