merge: resolve conflicts with develop (supersede by PR #89 / #107)

Three conflicts — all resolved in favor of develop, which has a more
refined version of the same work this branch introduced:

- radar_system_top.v: develop's cleaner USB_MODE=1 comment (same value).
- run_regression.sh: develop's ${SYSTEM_RTL[@]} refactor + added
  USB_MODE=1 test variants.
- tb/radar_system_tb.v: develop's ifdef USB_MODE_1 to dump the correct
  USB instance based on mode.

The 400 MHz reset fan-out fix (nco_400m_enhanced, cic_decimator_4x_enhanced,
ddc_400m) and ADAR1000 channel-indexing fix remain intact on this branch.

9_Firmware/9_1_Microcontroller/9_1_1_C_Cpp_Libraries/ADAR1000_Manager.cpp

@@ -868,11 +868,22 @@ void ADAR1000Manager::adarSetRamBypass(uint8_t deviceIndex, uint8_t broadcast) {
 }
 
 void ADAR1000Manager::adarSetRxPhase(uint8_t deviceIndex, uint8_t channel, uint8_t phase, uint8_t broadcast) {
+    // channel is 1-based (CH1..CH4) per API contract documented in
+    // ADAR1000_AGC.cpp and matching ADI datasheet terminology.
+    // Reject out-of-range early so a stale 0-based caller does not
+    // silently wrap to ((0-1) & 0x03) == 3 and write to CH4.
+    // See issue #90.
+    if (channel < 1 || channel > 4) {
+        DIAG("BF", "adarSetRxPhase: channel %u out of range [1..4], ignored", channel);
+        return;
+    }
     uint8_t i_val = VM_I[phase % 128];
     uint8_t q_val = VM_Q[phase % 128];
 
-    uint32_t mem_addr_i = REG_CH1_RX_PHS_I + (channel & 0x03) * 2;
+    // Subtract 1 to convert 1-based channel to 0-based register offset
-    uint32_t mem_addr_q = REG_CH1_RX_PHS_Q + (channel & 0x03) * 2;
+    // before masking. See issue #90.
+    uint32_t mem_addr_i = REG_CH1_RX_PHS_I + ((channel - 1) & 0x03) * 2;
+    uint32_t mem_addr_q = REG_CH1_RX_PHS_Q + ((channel - 1) & 0x03) * 2;
 
     adarWrite(deviceIndex, mem_addr_i, i_val, broadcast);
     adarWrite(deviceIndex, mem_addr_q, q_val, broadcast);
@@ -880,11 +891,16 @@ void ADAR1000Manager::adarSetRxPhase(uint8_t deviceIndex, uint8_t channel, uint8
 }
 
 void ADAR1000Manager::adarSetTxPhase(uint8_t deviceIndex, uint8_t channel, uint8_t phase, uint8_t broadcast) {
+    // channel is 1-based (CH1..CH4). See issue #90.
+    if (channel < 1 || channel > 4) {
+        DIAG("BF", "adarSetTxPhase: channel %u out of range [1..4], ignored", channel);
+        return;
+    }
     uint8_t i_val = VM_I[phase % 128];
     uint8_t q_val = VM_Q[phase % 128];
 
-    uint32_t mem_addr_i = REG_CH1_TX_PHS_I + (channel & 0x03) * 2;
+    uint32_t mem_addr_i = REG_CH1_TX_PHS_I + ((channel - 1) & 0x03) * 2;
-    uint32_t mem_addr_q = REG_CH1_TX_PHS_Q + (channel & 0x03) * 2;
+    uint32_t mem_addr_q = REG_CH1_TX_PHS_Q + ((channel - 1) & 0x03) * 2;
 
     adarWrite(deviceIndex, mem_addr_i, i_val, broadcast);
     adarWrite(deviceIndex, mem_addr_q, q_val, broadcast);
@@ -892,13 +908,23 @@ void ADAR1000Manager::adarSetTxPhase(uint8_t deviceIndex, uint8_t channel, uint8
 }
 
 void ADAR1000Manager::adarSetRxVgaGain(uint8_t deviceIndex, uint8_t channel, uint8_t gain, uint8_t broadcast) {
-    uint32_t mem_addr = REG_CH1_RX_GAIN + (channel & 0x03);
+    // channel is 1-based (CH1..CH4). See issue #90.
+    if (channel < 1 || channel > 4) {
+        DIAG("BF", "adarSetRxVgaGain: channel %u out of range [1..4], ignored", channel);
+        return;
+    }
+    uint32_t mem_addr = REG_CH1_RX_GAIN + ((channel - 1) & 0x03);
     adarWrite(deviceIndex, mem_addr, gain, broadcast);
     adarWrite(deviceIndex, REG_LOAD_WORKING, 0x1, broadcast);
 }
 
 void ADAR1000Manager::adarSetTxVgaGain(uint8_t deviceIndex, uint8_t channel, uint8_t gain, uint8_t broadcast) {
-    uint32_t mem_addr = REG_CH1_TX_GAIN + (channel & 0x03);
+    // channel is 1-based (CH1..CH4). See issue #90.
+    if (channel < 1 || channel > 4) {
+        DIAG("BF", "adarSetTxVgaGain: channel %u out of range [1..4], ignored", channel);
+        return;
+    }
+    uint32_t mem_addr = REG_CH1_TX_GAIN + ((channel - 1) & 0x03);
     adarWrite(deviceIndex, mem_addr, gain, broadcast);
     adarWrite(deviceIndex, REG_LOAD_WORKING, LD_WRK_REGS_LDTX_OVERRIDE, broadcast);
 }

9_Firmware/9_2_FPGA/cic_decimator_4x_enhanced.v

@@ -32,11 +32,50 @@ localparam COMB_WIDTH = 28;
 // adjacent DSP48E1 tiles — zero fabric delay, guaranteed to meet 400+ MHz
 // on 7-series regardless of speed grade.
 //
-// Active-high reset derived from reset_n (inverted).
+// Active-high reset derived from reset_n (inverted and REGISTERED).
 // CEP (clock enable for P register) gated by data_valid.
-// ============================================================================
+//
-
+// ----------------------------------------------------------------------------
-wire reset_h = ~reset_n;  // active-high reset for DSP48E1 RSTP
+// RESET FAN-OUT INVARIANT (Build N+1 fix for WNS=-0.626ns at 400 MHz):
+// ----------------------------------------------------------------------------
+// Previously this was a combinational wire (`wire reset_h = ~reset_n`). Vivado
+// collapsed all per-module inversions across the DDC hierarchy into a SINGLE
+// shared LUT1, whose output fanned out to 702 loads (DSP48E1 RSTP/RSTB/RSTC
+// plus FDRE R pins of all comb-stage DSP48E1s inferred via use_dsp="yes").
+// Route delay alone on that net was 2.019–2.268 ns — nearly one full 2.5 ns
+// period. Timing failed by 626 ps on the 400 MHz domain.
+//
+// Fix: convert reset_h to a REGISTERED signal with (* max_fanout = 50 *).
+// Vivado treats max_fanout on a REG (not a wire) as authoritative and
+// replicates the register into N copies, each placed near its ≈50 loads.
+// Invariants preserved:
+//   I1 (correctness):       reset_h is still active-high, equals ~reset_n
+//                           after one clk edge; CIC reset is a RECEIVER-side
+//                           synchronizer anyway (driven by reset_n_400m which
+//                           is already sync'd in the parent DDC), so adding
+//                           one more clk cycle of latency is safe.
+//   I2 (glitch-free):       Registered output => inherently glitch-free,
+//                           feeding DSP48E1 RST pins (which are synchronous
+//                           to CLK, so they capture on the same edge anyway).
+//   I3 (power-up safety):   reset_h is NOT async-reset itself. On power-up,
+//                           FDRE INIT=0 starts reset_h LOW. First clk edge
+//                           samples ~reset_n which is LOW on power-up (the
+//                           parent DDC holds reset_n_400m low until the 2-
+//                           stage synchronizer releases), so reset_h goes
+//                           HIGH on cycle 1 and all DSPs see reset during
+//                           the following cycles. System is held in reset
+//                           for enough cycles that any initial register
+//                           state garbage is overwritten. ✅
+//   I4 (reset de-assertion):reset_h goes LOW one cycle AFTER reset_n_400m
+//                           goes HIGH. Downstream DSPs come out of reset on
+//                           the next clk edge after that. Total latency
+//                           from system reset release to first valid sample:
+//                           2 (sync chain) + 1 (reset_h reg) + 1 (first
+//                           DSP output) = 4 cycles at 400 MHz = 10 ns.
+//                           Negligible vs system reset assertion duration.
+// ----------------------------------------------------------------------------
+(* max_fanout = 50 *) reg reset_h = 1'b1;  // INIT=1'b1: registers start in reset state on power-up
+always @(posedge clk) reset_h <= ~reset_n;
 
 // Sign-extended input for integrator_0 C port (48-bit)
 wire [ACC_WIDTH-1:0] data_in_c = {{(ACC_WIDTH-18){data_in[17]}}, data_in};
@@ -699,10 +738,11 @@ initial begin
 end
 
 // Decimation control + monitoring (integrators are now DSP48E1 instances)
-// Sync reset: enables FDRE inference for better timing at 400 MHz.
+// Sync reset via reset_h (registered, max_fanout=50) — eliminates the shared
-// Reset is already synchronous to clk via reset synchronizer in parent module.
+// LUT1 inverter that previously fanned out to all fabric FDRE R pins plus
+// DSP48E1 RST pins (702 loads total). See "RESET FAN-OUT INVARIANT" at top.
 always @(posedge clk) begin
-    if (!reset_n) begin
+    if (reset_h) begin
         integrator_sampled <= 0;
         decimation_counter <= 0;
         data_valid_delayed <= 0;
@@ -755,9 +795,9 @@ always @(posedge clk) begin
 end
 
 // Pipeline the valid signal for comb section
-// Sync reset: matches decimation control block reset style.
+// Sync reset via reset_h — same replicated-register source as DSP48E1 RSTs.
 always @(posedge clk) begin
-    if (!reset_n) begin
+    if (reset_h) begin
         data_valid_comb <= 0;
         data_valid_comb_pipe <= 0;
         data_valid_comb_0_out <= 0;
@@ -792,7 +832,7 @@ end
 //   - Each stage: comb[i] = comb[i-1] - comb_delay[i][last]
 
 always @(posedge clk) begin
-    if (!reset_n) begin
+    if (reset_h) begin
         for (i = 0; i < STAGES; i = i + 1) begin
             comb[i] <= 0;
             for (j = 0; j < COMB_DELAY; j = j + 1) begin

9_Firmware/9_2_FPGA/ddc_400m.v

@@ -53,46 +53,6 @@ reg [2:0] saturation_count;
 reg overflow_detected;
 reg [7:0] error_counter;
 
-// ============================================================================
-// 400 MHz Reset Synchronizer
-//
-// reset_n arrives from the 100 MHz domain (sys_reset_n from radar_system_top).
-// Using it directly as an async reset in the 400 MHz domain causes the reset
-// deassertion edge to violate timing: the 100 MHz flip-flop driving reset_n
-// has its output fanning out to 1156 registers across the FPGA in the 400 MHz
-// domain, requiring 18.243ns of routing (WNS = -18.081ns).
-//
-// Solution: 2-stage async-assert, sync-deassert reset synchronizer in the
-// 400 MHz domain. Reset assertion is immediate (asynchronous — combinatorial
-// path from reset_n to all 400 MHz registers). Reset deassertion is
-// synchronized to clk_400m rising edge, preventing metastability.
-//
-// All 400 MHz submodules (NCO, CIC, mixers, LFSR) use reset_n_400m.
-// All 100 MHz submodules (FIR, output stage) continue using reset_n directly
-// (already synchronized to 100 MHz at radar_system_top level).
-// ============================================================================
-(* ASYNC_REG = "TRUE" *) reg [1:0] reset_sync_400m;
-(* max_fanout = 50 *) wire reset_n_400m = reset_sync_400m[1];
-
-// Active-high reset for DSP48E1 RST ports (avoids LUT1 inverter fan-out)
-(* max_fanout = 50 *) reg reset_400m;
-
-always @(posedge clk_400m or negedge reset_n) begin
-    if (!reset_n) begin
-        reset_sync_400m <= 2'b00;
-        reset_400m      <= 1'b1;
-    end else begin
-        reset_sync_400m <= {reset_sync_400m[0], 1'b1};
-        reset_400m      <= ~reset_sync_400m[1];
-    end
-end
-
-// CDC synchronization for control signals (2-stage synchronizers)
-(* ASYNC_REG = "TRUE" *) reg [1:0] mixers_enable_sync_chain;
-(* ASYNC_REG = "TRUE" *) reg [1:0] force_saturation_sync_chain;
-wire mixers_enable_sync;
-wire force_saturation_sync;
-
 // Debug monitoring signals
 reg [31:0] sample_counter;
 wire signed [17:0] debug_mixed_i_trunc;
@@ -130,8 +90,6 @@ reg baseband_valid_reg;
 wire [7:0] phase_dither_bits;
 reg [31:0] phase_inc_dithered;
 
-
-
 // ============================================================================
 // Debug Signal Assignments
 // ============================================================================
@@ -142,13 +100,66 @@ assign debug_mixed_i_trunc = mixed_i[25:8];
 assign debug_mixed_q_trunc = mixed_q[25:8];
 
 // ============================================================================
-// Clock Domain Crossing for Control Signals (2-stage synchronizers)
+// 400 MHz Reset Synchronizer
+//
+// reset_n arrives from the 100 MHz domain (sys_reset_n from radar_system_top).
+// Using it directly as an async reset in the 400 MHz domain causes the reset
+// deassertion edge to violate timing: the 100 MHz flip-flop driving reset_n
+// has its output fanning out to 1156 registers across the FPGA in the 400 MHz
+// domain, requiring 18.243ns of routing (WNS = -18.081ns).
+//
+// Solution: 2-stage async-assert, sync-deassert reset synchronizer in the
+// 400 MHz domain. Reset assertion is immediate (asynchronous — combinatorial
+// path from reset_n to all 400 MHz registers). Reset deassertion is
+//
+// reset_400m   : ACTIVE-HIGH registered reset with (* max_fanout = 50 *).
+//                This is THE signal fed to every synchronous 400 MHz FDRE
+//                and every DSP48E1 RST pin in this module and its children
+//                (NCO, CIC, LFSR). Vivado replicates the register (~14
+//                copies) so each replica drives ≈50 loads regionally,
+//                eliminating the single-LUT1 / 702-load net that caused
+//                WNS=-0.626 ns in Build N.
+//
+// System-level invariants preserved:
+//   I1  Reset assertion propagates to all 400 MHz regs within ≤3 clk edges
+//       (2 sync + 1 replicated-reg fanout).  At 400 MHz = 7.5 ns << any
+//       system-level reset assertion duration.
+//   I2  Reset de-assertion is always synchronous to clk_400m (via
+//       reset_sync_400m), never glitches.
+//   I3  DSP48E1 RST pins are all fed from Q of a register — glitch-free.
+//   I4  No new CDC introduced: reset_400m is entirely in clk_400m domain.
+//   I5  Power-up: reset_n is asserted externally and mmcm_locked is low;
+//       reset_sync_400m stays 2'b00, reset_400m stays 1'b1, downstream
+//       FDREs stay cleared. Safe.
 // ============================================================================
+(* ASYNC_REG = "TRUE" *) reg [1:0] reset_sync_400m = 2'b00;
+(* max_fanout = 50 *) wire reset_n_400m = reset_sync_400m[1];
+
+// Active-high replicated reset for all synchronous 400 MHz consumers
+(* max_fanout = 50 *) reg reset_400m = 1'b1;
+
+always @(posedge clk_400m or negedge reset_n) begin
+    if (!reset_n) begin
+        reset_sync_400m <= 2'b00;
+        reset_400m      <= 1'b1;
+    end else begin
+        reset_sync_400m <= {reset_sync_400m[0], 1'b1};
+        reset_400m      <= ~reset_sync_400m[1];
+    end
+end
+
+// CDC synchronization for control signals (2-stage synchronizers)
+(* ASYNC_REG = "TRUE" *) reg [1:0] mixers_enable_sync_chain;
+(* ASYNC_REG = "TRUE" *) reg [1:0] force_saturation_sync_chain;
+wire mixers_enable_sync;
+wire force_saturation_sync;
 assign mixers_enable_sync = mixers_enable_sync_chain[1];
 assign force_saturation_sync = force_saturation_sync_chain[1];
 
-always @(posedge clk_400m or negedge reset_n_400m) begin
+// Sync reset via reset_400m (replicated, max_fanout=50). Was async on
-    if (!reset_n_400m) begin
+// reset_n_400m — see "400 MHz RESET DISTRIBUTION" comment above.
+always @(posedge clk_400m) begin
+    if (reset_400m) begin
         mixers_enable_sync_chain <= 2'b00;
         force_saturation_sync_chain <= 2'b00;
     end else begin
@@ -160,8 +171,8 @@ end
 // ============================================================================
 // Sample Counter and Debug Monitoring
 // ============================================================================
-always @(posedge clk_400m or negedge reset_n_400m) begin
+always @(posedge clk_400m) begin
-    if (!reset_n_400m || reset_monitors) begin
+    if (reset_400m || reset_monitors) begin
         sample_counter <= 0;
         error_counter <= 0;
     end else if (adc_data_valid_i && adc_data_valid_q ) begin
@@ -189,8 +200,8 @@ lfsr_dither_enhanced #(
 localparam PHASE_INC_120MHZ = 32'h4CCCCCCD;
 
 // Apply dithering to reduce spurious tones (registered for 400 MHz timing)
-always @(posedge clk_400m or negedge reset_n_400m) begin
+always @(posedge clk_400m) begin
-    if (!reset_n_400m)
+    if (reset_400m)
         phase_inc_dithered <= PHASE_INC_120MHZ;
     else
         phase_inc_dithered <= PHASE_INC_120MHZ + {24'b0, phase_dither_bits};
@@ -229,8 +240,8 @@ assign adc_signed_w = {1'b0, adc_data, {(MIXER_WIDTH-ADC_WIDTH-1){1'b0}}} -
                       {1'b0, {ADC_WIDTH{1'b1}}, {(MIXER_WIDTH-ADC_WIDTH-1){1'b0}}} / 2;
 
 // Valid pipeline: 5-stage shift register (1 NCO pipe + 3 DSP48E1 AREG+MREG+PREG + 1 retiming)
-always @(posedge clk_400m or negedge reset_n_400m) begin
+always @(posedge clk_400m) begin
-    if (!reset_n_400m) begin
+    if (reset_400m) begin
         dsp_valid_pipe <= 5'b00000;
     end else begin
         dsp_valid_pipe <= {dsp_valid_pipe[3:0], (nco_ready && adc_data_valid_i && adc_data_valid_q)};
@@ -246,8 +257,8 @@ reg signed [MIXER_WIDTH+NCO_WIDTH-1:0] mult_i_internal, mult_q_internal;  // Mod
 reg signed [MIXER_WIDTH+NCO_WIDTH-1:0] mult_i_reg, mult_q_reg;            // Models PREG
 
 // Stage 0: NCO pipeline — breaks long NCO→DSP route (matches synthesis fabric registers)
-always @(posedge clk_400m or negedge reset_n_400m) begin
+always @(posedge clk_400m) begin
-    if (!reset_n_400m) begin
+    if (reset_400m) begin
         cos_nco_pipe <= 0;
         sin_nco_pipe <= 0;
     end else begin
@@ -257,8 +268,8 @@ always @(posedge clk_400m or negedge reset_n_400m) begin
 end
 
 // Stage 1: AREG/BREG equivalent (uses pipelined NCO outputs)
-always @(posedge clk_400m or negedge reset_n_400m) begin
+always @(posedge clk_400m) begin
-    if (!reset_n_400m) begin
+    if (reset_400m) begin
         adc_signed_reg <= 0;
         cos_pipe_reg <= 0;
         sin_pipe_reg <= 0;
@@ -270,8 +281,8 @@ always @(posedge clk_400m or negedge reset_n_400m) begin
 end
 
 // Stage 2: MREG equivalent
-always @(posedge clk_400m or negedge reset_n_400m) begin
+always @(posedge clk_400m) begin
-    if (!reset_n_400m) begin
+    if (reset_400m) begin
         mult_i_internal <= 0;
         mult_q_internal <= 0;
     end else begin
@@ -281,8 +292,8 @@ always @(posedge clk_400m or negedge reset_n_400m) begin
 end
 
 // Stage 3: PREG equivalent
-always @(posedge clk_400m or negedge reset_n_400m) begin
+always @(posedge clk_400m) begin
-    if (!reset_n_400m) begin
+    if (reset_400m) begin
         mult_i_reg <= 0;
         mult_q_reg <= 0;
     end else begin
@@ -292,8 +303,8 @@ always @(posedge clk_400m or negedge reset_n_400m) begin
 end
 
 // Stage 4: Post-DSP retiming register (matches synthesis path)
-always @(posedge clk_400m or negedge reset_n_400m) begin
+always @(posedge clk_400m) begin
-    if (!reset_n_400m) begin
+    if (reset_400m) begin
         mult_i_retimed <= 0;
         mult_q_retimed <= 0;
     end else begin
@@ -311,8 +322,8 @@ wire [47:0] dsp_p_i, dsp_p_q;
 // (1.505ns routing observed in Build 26). These fabric registers are placed
 // near the DSP by the placer, splitting the route into two shorter segments.
 // DONT_TOUCH on the reg declaration (above) prevents absorption/retiming.
-always @(posedge clk_400m or negedge reset_n_400m) begin
+always @(posedge clk_400m) begin
-    if (!reset_n_400m) begin
+    if (reset_400m) begin
         cos_nco_pipe <= 0;
         sin_nco_pipe <= 0;
     end else begin
@@ -329,11 +340,10 @@ DSP48E1 #(
     .USE_DPORT("FALSE"),
     .USE_MULT("MULTIPLY"),
     .USE_SIMD("ONE48"),
-    // Pipeline register attributes — all enabled for max timing
     .AREG(1),
     .BREG(1),
     .MREG(1),
-    .PREG(1),           // P register enabled — absorbs CLK→P delay for timing closure
+    .PREG(1),
     .ADREG(0),
     .ACASCREG(1),
     .BCASCREG(1),
@@ -344,7 +354,6 @@ DSP48E1 #(
     .DREG(0),
     .INMODEREG(0),
     .OPMODEREG(0),
-    // Pattern detector (unused)
     .AUTORESET_PATDET("NO_RESET"),
     .MASK(48'h3fffffffffff),
     .PATTERN(48'h000000000000),
@@ -496,8 +505,8 @@ wire signed [MIXER_WIDTH+NCO_WIDTH-1:0] mult_q_reg = dsp_p_q[MIXER_WIDTH+NCO_WID
 // Stage 4: Post-DSP retiming register — breaks DSP48E1 CLK→P to fabric path
 // Without this, the DSP output prop delay (1.866ns) + routing (0.515ns) exceeds
 // the 2.500ns clock period at slow process corner
-always @(posedge clk_400m or negedge reset_n_400m) begin
+always @(posedge clk_400m) begin
-    if (!reset_n_400m) begin
+    if (reset_400m) begin
         mult_i_retimed <= 0;
         mult_q_retimed <= 0;
     end else begin
@@ -513,8 +522,8 @@ end
 // force_saturation mux is intentionally AFTER the DSP48E1 output to avoid
 // polluting the critical input path with extra logic
 // ============================================================================
-always @(posedge clk_400m or negedge reset_n_400m) begin
+always @(posedge clk_400m) begin
-    if (!reset_n_400m) begin
+    if (reset_400m) begin
         mixed_i <= 0;
         mixed_q <= 0;
         mixed_valid <= 0;
@@ -759,8 +768,17 @@ generate
     end
 endgenerate
 
-always @(posedge clk or negedge reset_n) begin
+// ============================================================================
-    if (!reset_n) begin
+// RESET FAN-OUT INVARIANT: registered active-high reset with max_fanout=50.
+// See cic_decimator_4x_enhanced.v for full reasoning. reset_n here is driven
+// by the parent DDC's reset_n_400m (already synchronized to clk_400m), so
+// sync reset on the LFSR is safe. INIT=1'b1 holds LFSR in reset on power-up.
+// ============================================================================
+(* max_fanout = 50 *) reg reset_h = 1'b1;
+always @(posedge clk) reset_h <= ~reset_n;
+
+always @(posedge clk) begin
+    if (reset_h) begin
         lfsr_reg <= {DITHER_WIDTH{1'b1}};  // Non-zero initial state
         cycle_counter <= 0;
         lock_detected <= 0;

9_Firmware/9_2_FPGA/nco_400m_enhanced.v

@@ -59,6 +59,25 @@ reg [1:0] quadrant_reg2;                 // Pass-through for Stage 5 MUX
 // Valid pipeline: tracks 6-stage latency
 reg [5:0] valid_pipe;
 
+// ============================================================================
+// RESET FAN-OUT INVARIANT (Build N+1 fix for WNS=-0.626ns at 400 MHz):
+// ============================================================================
+// reset_h is an ACTIVE-HIGH, REGISTERED copy of ~reset_n with (* max_fanout=50 *).
+// Vivado replicates this register (14+ copies) so each copy drives ≈50 loads
+// regionally, avoiding the single-LUT1 / 702-load net that caused timing
+// failure in Build N. It feeds:
+//   - DSP48E1 RSTP/RSTC on the phase-accumulator DSP (below)
+//   - All pipeline-stage fabric FDREs (synchronous reset)
+// Invariants (see cic_decimator_4x_enhanced.v for full reasoning):
+//   I1 correctness:      reset_h == ~reset_n one cycle later
+//   I2 glitch-free:      registered output
+//   I3 power-up safe:    INIT=1'b1 holds all downstream in reset until first
+//                        valid clock edge; reset_n is low on power-up anyway
+//   I4 de-assert lat.:   +1 cycle vs. direct async; negligible at 400 MHz
+// ============================================================================
+(* max_fanout = 50 *) reg reset_h = 1'b1;
+always @(posedge clk_400m) reset_h <= ~reset_n;
+
 // Use only the top 8 bits for LUT addressing (256-entry LUT equivalent)
 wire [7:0] lut_address = phase_with_offset[31:24];
 
@@ -135,8 +154,8 @@ wire [15:0] cos_abs_w = sin_lut[63 - lut_index_pipe_cos];
 // Stage 2: phase_with_offset adds phase offset
 reg [31:0] phase_accumulator;
 
-always @(posedge clk_400m or negedge reset_n) begin
+always @(posedge clk_400m) begin
-    if (!reset_n) begin
+    if (reset_h) begin
         phase_accumulator <= 32'h00000000;
         phase_accum_reg   <= 32'h00000000;
         phase_with_offset <= 32'h00000000;
@@ -190,8 +209,8 @@ DSP48E1 #(
     .RSTA(1'b0),
     .RSTB(1'b0),
     .RSTM(1'b0),
-    .RSTP(!reset_n),             // Reset P register (phase accumulator) on !reset_n
+    .RSTP(reset_h),              // Reset P register (phase accumulator) — registered, max_fanout=50
-    .RSTC(!reset_n),             // Reset C register (tuning word) on !reset_n
+    .RSTC(reset_h),              // Reset C register (tuning word)       — registered, max_fanout=50
     .RSTALLCARRYIN(1'b0),
     .RSTALUMODE(1'b0),
     .RSTCTRL(1'b0),
@@ -245,8 +264,8 @@ DSP48E1 #(
 // Stage 1: Capture DSP48E1 P output into fabric register
 // Stage 2: Add phase offset to captured value
 // Split into two registered stages to break DSP48E1.P→CARRY4 critical path
-always @(posedge clk_400m or negedge reset_n) begin
+always @(posedge clk_400m) begin
-    if (!reset_n) begin
+    if (reset_h) begin
         phase_accum_reg   <= 32'h00000000;
         phase_with_offset <= 32'h00000000;
     end else if (phase_valid) begin
@@ -264,8 +283,8 @@ end
 //           Only 2 registers driven (lut_index_pipe + quadrant_pipe)
 //           Minimal fanout → short routes → easy timing
 // ============================================================================
-always @(posedge clk_400m or negedge reset_n) begin
+always @(posedge clk_400m) begin
-    if (!reset_n) begin
+    if (reset_h) begin
         lut_index_pipe_sin <= 6'b000000;
         lut_index_pipe_cos <= 6'b000000;
         quadrant_pipe      <= 2'b00;
@@ -281,8 +300,8 @@ end
 //           Registered address → combinational LUT6 read → register
 //           Only 1 logic level (LUT6), trivial timing
 // ============================================================================
-always @(posedge clk_400m or negedge reset_n) begin
+always @(posedge clk_400m) begin
-    if (!reset_n) begin
+    if (reset_h) begin
         sin_abs_reg <= 16'h0000;
         cos_abs_reg <= 16'h7FFF;
         quadrant_reg <= 2'b00;
@@ -298,8 +317,8 @@ end
 //          CARRY4 x4 chain has registered inputs — easily fits in 2.5ns
 //          Also pass through abs values and quadrant for Stage 5
 // ============================================================================
-always @(posedge clk_400m or negedge reset_n) begin
+always @(posedge clk_400m) begin
-    if (!reset_n) begin
+    if (reset_h) begin
         sin_neg_reg <= 16'h0000;
         cos_neg_reg <= -16'h7FFF;
         sin_abs_reg2 <= 16'h0000;
@@ -318,8 +337,8 @@ end
 // Stage 5: Quadrant sign application → final sin/cos output
 //          Uses pre-computed negated values from Stage 4 — pure MUX, no arithmetic
 // ============================================================================
-always @(posedge clk_400m or negedge reset_n) begin
+always @(posedge clk_400m) begin
-    if (!reset_n) begin
+    if (reset_h) begin
         sin_out <= 16'h0000;
         cos_out <= 16'h7FFF;
     end else if (valid_pipe[4]) begin
@@ -347,8 +366,8 @@ end
 // ============================================================================
 // Valid pipeline and dds_ready (6-stage latency)
 // ============================================================================
-always @(posedge clk_400m or negedge reset_n) begin
+always @(posedge clk_400m) begin
-    if (!reset_n) begin
+    if (reset_h) begin
         valid_pipe <= 6'b000000;
         dds_ready <= 1'b0;
     end else begin

9_Firmware/tests/cross_layer/test_cross_layer_contract.py

@@ -26,12 +26,14 @@ layers agree (because both could be wrong).
 
 from __future__ import annotations
 
+import ast
 import os
 import re
 import struct
 import subprocess
 import tempfile
 from pathlib import Path
+from typing import ClassVar
 
 import pytest
 
@@ -625,6 +627,420 @@ class TestTier1AgcCrossLayerInvariant:
         )
 
 
+# ===================================================================
+# ADAR1000 channel→register round-trip invariant (issue #90)
+# ===================================================================
+#
+# Ground-truth invariant crossing three system layers:
+#   Chip (datasheet) -> Driver (MCU helpers) -> Application (callers).
+#
+# For every logical element ch in {0,1,2,3} (hardware channels CH1..CH4),
+# the round-trip
+#     caller_expr(ch) --> helper_offset(channel) * stride --> base + off
+# must land on the physical register REG_CH{ch+1}_* defined in the ADI
+# ADAR1000 register map parsed from ADAR1000_Manager.h.
+#
+# Catches:
+#   * #90 channel rotation regardless of which side is fixed (caller OR helper).
+#   * Wrong stride (e.g. phase written with stride 1 instead of 2).
+#   * Bad mask (e.g. `channel & 0x07`, `channel & 0x01`).
+#   * Wrong base register in a helper.
+#   * New setter added with mismatched convention.
+#   * Caller moved to a file the test no longer scans (fails loudly).
+#
+# Cannot be defeated by:
+#   * Renaming/refactoring helper layout: the setter coverage test
+#     (`test_helper_sites_exist_for_all_setters`) catches missing parse.
+#   * Changing 0x03 to 3 or adding a named constant: the offset is
+#     evaluated symbolically via AST, not matched by regex.
+
+
+def _parse_adar_register_map(header_text):
+    """Extract `#define REG_CHn_(RX|TX)_(GAIN|PHS_I|PHS_Q)` values."""
+    regs = {}
+    for m in re.finditer(
+        r"^#define\s+(REG_CH[1-4]_(?:RX|TX)_(?:GAIN|PHS_I|PHS_Q))\s+(0x[0-9A-Fa-f]+)",
+        header_text,
+        re.MULTILINE,
+    ):
+        regs[m.group(1)] = int(m.group(2), 16)
+    return regs
+
+
+def _safe_eval_int_expr(expr, **variables):
+    """
+    Evaluate a small integer expression with +, -, *, &, |, ^, ~, <<, >>.
+    Python's & / | / ^ / ~ / << / >> have the same semantics as C for the
+    operand widths we care about here (uint8_t after the mask makes the
+    result fit in 0..3). No floating point, no function calls, no names
+    outside ``variables``.
+
+    SECURITY: ``expr`` MUST come from a trusted source -- specifically,
+    C/C++ source text under version control in this repository (e.g.
+    arguments parsed out of ``main.cpp``/``ADAR1000_AGC.cpp``).  Although
+    the AST whitelist below rejects function calls, attribute access,
+    subscripts, and any name not in ``variables``, ``eval`` is still
+    invoked on the compiled tree.  Do NOT pass user-supplied / network /
+    GUI input here.
+    """
+    tree = ast.parse(expr, mode="eval")
+    allowed = (
+        ast.Expression, ast.BinOp, ast.UnaryOp, ast.Constant,
+        ast.Name, ast.Load,
+        ast.Add, ast.Sub, ast.Mult, ast.Mod, ast.FloorDiv,
+        ast.BitAnd, ast.BitOr, ast.BitXor,
+        ast.USub, ast.UAdd, ast.Invert,
+        ast.LShift, ast.RShift,
+    )
+    for node in ast.walk(tree):
+        if not isinstance(node, allowed):
+            raise ValueError(
+                f"disallowed AST node {type(node).__name__!s} in `{expr}`"
+            )
+    return eval(
+        compile(tree, "<expr>", "eval"),
+        {"__builtins__": {}},
+        variables,
+    )
+
+
+def _extract_adar_helper_sites(manager_cpp, setter_names):
+    """
+    For each setter, locate the body of ``void ADAR1000Manager::<setter>``
+    and return a list of (setter, base_register, offset_expr_c, stride)
+    for every ``REG_CHn_XXX + <expr>`` memory-address assignment.
+    """
+    sites = []
+    for setter in setter_names:
+        m = re.search(
+            rf"void\s+ADAR1000Manager::{setter}\s*\([^)]*\)\s*\{{(.+?)^\}}",
+            manager_cpp,
+            re.MULTILINE | re.DOTALL,
+        )
+        if not m:
+            continue
+        body = m.group(1)
+        for access in re.finditer(
+            r"=\s*(REG_CH[1-4]_(?:RX|TX)_(?:GAIN|PHS_I|PHS_Q))\s*\+\s*([^;]+);",
+            body,
+        ):
+            base = access.group(1)
+            rhs = access.group(2).strip()
+            # Trailing `* <integer>` = stride multiplier (2 for phase I/Q).
+            stride_match = re.match(r"(.+?)\s*\*\s*(\d+)\s*$", rhs)
+            if stride_match:
+                offset_expr = stride_match.group(1).strip()
+                stride = int(stride_match.group(2))
+            else:
+                offset_expr = rhs
+                stride = 1
+            sites.append((setter, base, offset_expr, stride))
+    return sites
+
+
+# Method-definition line pattern: `[qualifier...] <ret-type> <Class>::<setter>(`
+# Covers: plain `void X::f(`, `inline void X::f(`, `static bool X::f(`, etc.
+_DEFN_RE = re.compile(
+    r"^\s*(?:inline\s+|static\s+|virtual\s+|constexpr\s+|explicit\s+)*"
+    r"(?:void|bool|uint\w+|int\w*|auto)\s+\S+::\w+\s*\("
+)
+
+
+def _extract_adar_caller_sites(sources, setter):
+    """
+    Find every call ``<obj>.<setter>(dev, <channel_expr>, ...)`` across
+    ``sources = [(filename, text), ...]``. Returns (filename, line_no,
+    channel_expr) for each. Skips function declarations/definitions.
+
+    Arg list up to matching `)`: restricted to a single line. All existing
+    call sites fit on one line; a future multi-line refactor would drop
+    callers from the scan, which the round-trip test surfaces loudly via
+    `assert callers` (rather than silently missing a site).
+    """
+    out = []
+    call_re = re.compile(rf"\b{setter}\s*\(([^;]*?)\)\s*;")
+    for filename, text in sources:
+        for line_no, line in enumerate(text.splitlines(), start=1):
+            # Skip method definition / declaration lines.
+            if _DEFN_RE.match(line):
+                continue
+            cm = call_re.search(line)
+            if not cm:
+                continue
+            args = _split_top_level_commas(cm.group(1))
+            if len(args) < 2:
+                continue
+            channel_expr = args[1].strip()
+            out.append((filename, line_no, channel_expr))
+    return out
+
+
+def _split_top_level_commas(text):
+    """Split on commas that sit at paren-depth 0 (ignores nested calls)."""
+    parts, depth, cur = [], 0, []
+    for ch in text:
+        if ch == "(":
+            depth += 1
+            cur.append(ch)
+        elif ch == ")":
+            depth -= 1
+            cur.append(ch)
+        elif ch == "," and depth == 0:
+            parts.append("".join(cur))
+            cur = []
+        else:
+            cur.append(ch)
+    if cur:
+        parts.append("".join(cur))
+    return parts
+
+
+class TestTier1Adar1000ChannelRegisterRoundTrip:
+    """
+    Cross-layer round-trip: caller channel expr -> helper offset formula
+    -> physical register address must equal REG_CH{ch+1}_* for every
+    caller and every ch in {0,1,2,3}.
+
+    See module-level block comment above and upstream issue #90.
+    """
+
+    _SETTERS = (
+        "adarSetRxPhase",
+        "adarSetTxPhase",
+        "adarSetRxVgaGain",
+        "adarSetTxVgaGain",
+    )
+
+    # Register base -> stride override. Parsed values of stride are
+    # trusted; this table is the independent ground truth for cross-check.
+    _EXPECTED_STRIDE: ClassVar[dict[str, int]] = {
+        "REG_CH1_RX_GAIN": 1,
+        "REG_CH1_TX_GAIN": 1,
+        "REG_CH1_RX_PHS_I": 2,
+        "REG_CH1_RX_PHS_Q": 2,
+        "REG_CH1_TX_PHS_I": 2,
+        "REG_CH1_TX_PHS_Q": 2,
+    }
+
+    @classmethod
+    def setup_class(cls):
+        cls.header_txt = (cp.MCU_LIB_DIR / "ADAR1000_Manager.h").read_text()
+        cls.manager_txt = (cp.MCU_LIB_DIR / "ADAR1000_Manager.cpp").read_text()
+        cls.reg_map = _parse_adar_register_map(cls.header_txt)
+        cls.helper_sites = _extract_adar_helper_sites(
+            cls.manager_txt, cls._SETTERS,
+        )
+        # Auto-discover every C++ TU under the MCU tree so a new caller
+        # added to e.g. a future ``ADAR1000_Calibration.cpp`` cannot
+        # silently escape the round-trip check (issue #90 reviewer note).
+        # Exclude any path containing a ``tests`` segment so this test
+        # does not parse its own fixtures.  The resulting list is
+        # deterministic (sorted) for reproducible parametrization.
+        scanned = []
+        seen = set()
+        for root in (cp.MCU_LIB_DIR, cp.MCU_CODE_DIR):
+            for path in sorted(root.rglob("*.cpp")):
+                if "tests" in path.parts:
+                    continue
+                if path in seen:
+                    continue
+                seen.add(path)
+                scanned.append((path.name, path.read_text()))
+        cls.sources = scanned
+        # Sanity: the two TUs known to call ADAR1000 setters at the time
+        # of issue #90 must be in scope.  If a future refactor renames or
+        # moves them this assert fires loudly rather than silently
+        # passing an empty round-trip.
+        scanned_names = {n for (n, _) in scanned}
+        for required in ("ADAR1000_AGC.cpp", "main.cpp", "ADAR1000_Manager.cpp"):
+            assert required in scanned_names, (
+                f"Auto-discovery missed `{required}`; check MCU_LIB_DIR / "
+                f"MCU_CODE_DIR roots in contract_parser.py."
+            )
+
+    # ---------- Tier A: chip ground truth ----------------------------
+
+    def test_register_map_gain_stride_is_one_per_channel(self):
+        """Datasheet invariant: RX/TX VGA gain registers are 1 byte apart."""
+        for kind in ("RX_GAIN", "TX_GAIN"):
+            for n in range(1, 4):
+                delta = (
+                    self.reg_map[f"REG_CH{n+1}_{kind}"]
+                    - self.reg_map[f"REG_CH{n}_{kind}"]
+                )
+                assert delta == 1, (
+                    f"ADAR1000 register map invariant broken: "
+                    f"REG_CH{n+1}_{kind} - REG_CH{n}_{kind} = {delta}, "
+                    f"datasheet says 1. Either the header was mis-edited "
+                    f"or ADI released a part with a different map."
+                )
+
+    def test_register_map_phase_stride_is_two_per_channel(self):
+        """Datasheet invariant: phase I/Q pairs occupy 2 bytes per channel."""
+        for kind in ("RX_PHS_I", "RX_PHS_Q", "TX_PHS_I", "TX_PHS_Q"):
+            for n in range(1, 4):
+                delta = (
+                    self.reg_map[f"REG_CH{n+1}_{kind}"]
+                    - self.reg_map[f"REG_CH{n}_{kind}"]
+                )
+                assert delta == 2, (
+                    f"ADAR1000 register map invariant broken: "
+                    f"REG_CH{n+1}_{kind} - REG_CH{n}_{kind} = {delta}, "
+                    f"datasheet says 2."
+                )
+
+    # ---------- Tier B: driver parses cleanly -------------------------
+
+    def test_helper_sites_exist_for_all_setters(self):
+        """Every channel-indexed setter must parse at least one register access."""
+        found = {s for (s, _, _, _) in self.helper_sites}
+        missing = set(self._SETTERS) - found
+        assert not missing, (
+            f"Helper parse failed for: {sorted(missing)}. "
+            f"Either a setter was renamed (update _SETTERS), moved out of "
+            f"ADAR1000_Manager.cpp (extend scan scope), or the register-"
+            f"access form changed beyond `REG_CHn_XXX + <expr>`. "
+            f"DO NOT weaken this test without reviewing issue #90."
+        )
+
+    def test_helper_parsed_stride_matches_datasheet(self):
+        """Parsed helper strides must match the datasheet register spacing."""
+        for setter, base, offset_expr, stride in self.helper_sites:
+            expected = self._EXPECTED_STRIDE.get(base)
+            assert expected is not None, (
+                f"{setter} writes to unrecognised base `{base}`. "
+                f"If ADI added a new channel-indexed register block, "
+                f"extend _EXPECTED_STRIDE with its datasheet stride."
+            )
+            assert stride == expected, (
+                f"{setter} helper uses stride {stride} for `{base}` "
+                f"(`{offset_expr} * {stride}`), datasheet says {expected}. "
+                f"Writes will overlap or skip channels."
+            )
+
+    # ---------- Tier C: round-trip to physical register ---------------
+
+    def test_all_callers_pass_one_based_channel(self):
+        """
+        INVARIANT: every caller's channel argument must, for ch in
+        {0,1,2,3}, evaluate to a 1-based ADI channel index in {1,2,3,4}.
+
+        The bug fixed in #90 was that helpers used ``channel & 0x03``
+        directly, so a caller passing bare ``ch`` (0..3) appeared to
+        work for ch=0..2 and silently aliased ch=3 onto CH4-then-CH1.
+        After the fix, helpers do ``(channel - 1) & 0x03`` and reject
+        ``channel < 1 || channel > 4``.  A future caller written as
+        ``adarSetRxPhase(dev, ch, ...)`` (bare 0-based) or
+        ``adarSetRxPhase(dev, 0, ...)`` (literal 0) would silently be
+        dropped by the bounds-check at runtime; this test catches it at
+        CI time instead.
+
+        The check intentionally lives one tier above the round-trip test
+        so the failure message points the reader at the API contract
+        (1-based per ADI datasheet & ADAR1000_AGC.cpp:76) rather than at
+        a register-arithmetic mismatch.
+        """
+        offenders = []
+        for setter in self._SETTERS:
+            callers = _extract_adar_caller_sites(self.sources, setter)
+            for filename, line_no, ch_expr in callers:
+                for ch in range(4):
+                    try:
+                        channel_val = _safe_eval_int_expr(ch_expr, ch=ch)
+                    except (NameError, KeyError, ValueError) as e:
+                        offenders.append(
+                            f"  - {filename}:{line_no} {setter}("
+                            f"…, `{ch_expr}`, …) -- ch={ch}: "
+                            f"unparseable ({e})"
+                        )
+                        continue
+                    if channel_val not in (1, 2, 3, 4):
+                        offenders.append(
+                            f"  - {filename}:{line_no} {setter}("
+                            f"…, `{ch_expr}`, …) -- ch={ch}: "
+                            f"channel={channel_val}, expected 1..4"
+                        )
+        assert not offenders, (
+            "ADAR1000 1-based channel API contract violated. The fix "
+            "for issue #90 requires every caller to pass channel in "
+            "{1,2,3,4} (CH1..CH4 per ADI datasheet). Bare 0-based ch "
+            "or a literal 0 will be silently dropped by the helper's "
+            "bounds check.  Offenders:\n" + "\n".join(offenders)
+        )
+
+    @pytest.mark.parametrize(
+        "setter",
+        [
+            "adarSetRxPhase",
+            "adarSetTxPhase",
+            "adarSetRxVgaGain",
+            "adarSetTxVgaGain",
+        ],
+    )
+    def test_round_trip_lands_on_intended_physical_channel(self, setter):
+        """
+        INVARIANT: for every caller of ``<setter>`` and every logical ch
+        in {0,1,2,3}, the effective register address equals
+        REG_CH{ch+1}_*. Catches #90 regardless of fix direction.
+        """
+        callers = _extract_adar_caller_sites(self.sources, setter)
+        assert callers, (
+            f"No callers of `{setter}` found. Either the test scope is "
+            f"incomplete (extend `setup_class.sources`) or the symbol was "
+            f"inlined/removed. A blind test is a dangerous test — "
+            f"investigate before weakening."
+        )
+        helpers = [
+            (b, e, s) for (nm, b, e, s) in self.helper_sites if nm == setter
+        ]
+        assert helpers, f"helper body for `{setter}` not parseable"
+
+        errors = []
+        for filename, line_no, ch_expr in callers:
+            for ch in range(4):
+                try:
+                    channel_val = _safe_eval_int_expr(ch_expr, ch=ch)
+                except (NameError, KeyError, ValueError) as e:
+                    pytest.fail(
+                        f"{filename}:{line_no}: caller channel expression "
+                        f"`{ch_expr}` uses symbol outside {{ch}} or a "
+                        f"disallowed operator ({e}). Extend "
+                        f"_safe_eval_int_expr variables or rewrite the "
+                        f"call site with a supported expression."
+                    )
+                for base_sym, offset_expr, stride in helpers:
+                    try:
+                        offset = _safe_eval_int_expr(
+                            offset_expr, channel=channel_val,
+                        )
+                    except (NameError, KeyError, ValueError) as e:
+                        pytest.fail(
+                            f"helper `{setter}` offset expr "
+                            f"`{offset_expr}` uses symbol outside "
+                            f"{{channel}} or a disallowed operator ({e}). "
+                            f"Extend _safe_eval_int_expr variables if new "
+                            f"driver state is introduced."
+                        )
+                    final = self.reg_map[base_sym] + offset * stride
+                    expected_sym = base_sym.replace("CH1", f"CH{ch + 1}")
+                    expected = self.reg_map[expected_sym]
+                    if final != expected:
+                        errors.append(
+                            f"  - {filename}:{line_no} {setter} "
+                            f"caller `{ch_expr}` | ch={ch} -> "
+                            f"channel={channel_val} -> "
+                            f"`{base_sym} + ({offset_expr})"
+                            f"{' * ' + str(stride) if stride != 1 else ''}`"
+                            f" = 0x{final:03X} "
+                            f"(expected {expected_sym} = 0x{expected:03X})"
+                        )
+        assert not errors, (
+            f"ADAR1000 channel round-trip FAILED for {setter} "
+            f"({len(errors)} mismatches) — writes routed to wrong physical "
+            f"channel. This is issue #90.\n" + "\n".join(errors)
+        )
+
+
 class TestTier1DataPacketLayout:
     """Verify data packet byte layout matches between Python and Verilog."""