Add WAN link speed override for Adaptive SQM (#565)

* Add WAN link speed override for Adaptive SQM

Allow users to override the auto-detected WAN link speed, which is used
as the shaping ceiling (with 2% HTB headroom). This is needed when the
physical SFP supports a higher speed than what UniFi/ethtool reports
(e.g., 2.5G SFP in a UXG-Fiber reporting as 1G).

The override is nullable - when unset, behavior is identical to before
(auto-detected speed from the gateway port). The detected value shows
as placeholder text so the user knows what was auto-detected.

* Move Speed Margin to end of Calculated SQM Parameters grid

* Comment out Smart Queue download rate pre-deploy check

UniFi Network never clobbers our TC config once Adaptive SQM takes
over - the qdiscs get fully replaced by our scripts. The rate shown
in the UniFi UI becomes irrelevant after deployment.

* Clear link speed override when switching WAN interfaces

* Fix link speed override not persisting on update

The repository's upsert method copies properties individually for
existing rows. LinkSpeedOverrideMbps was missing from the update path,
so it was always written as null on save.

* Base speedtest probe rate on max shaping rate instead of link speed

The probe rate was 5% above max(AbsoluteMax, LinkSpeed), which could
overshoot the physical line rate and cause packet loss at the NIC level
where there is no AQM. Now it is 3% above MaxDownloadSpeed - the
highest rate the system will ever shape to - which keeps it just above
the shaper ceiling without exceeding what the wire can drain.

* Restore 5 KB burst tuning for HTB classes

5 KB burst eliminates downstream drop_overmemory for bulk flows at gig
speeds. Was reverted to stock 1500b pending per-connection testing but
the 2.5G link speed override makes this viable again.

* Restore original scaled burst function (rate * 5, clamped 1500-5000)

Author: tvancott42

src/NetworkOptimizer.Sqm/Models/SqmConfiguration.cs

@@ -66,11 +66,11 @@ public class SqmConfiguration
 
     /// <summary>
     /// Rate to set on TC during the speedtest probe so the measurement runs unshaped.
-    /// 5% above the larger of AbsoluteMax or physical link speed - safely above anything
-    /// the connection can actually deliver, so TC never constrains the test.
+    /// 3% above MaxDownloadSpeed (the highest rate the system will ever shape to).
+    /// Just enough headroom that TC stays transparent without overshooting physical line rate.
     /// </summary>
     public int SpeedtestProbeRateMbps =>
-        Math.Max(100, (int)(Math.Max(AbsoluteMaxDownloadSpeed, WanLinkSpeedMbps ?? 0) * 1.05));
+        Math.Max(100, (int)(MaxDownloadSpeed * 1.03));
 
     /// <summary>
     /// Overhead multiplier for speedtest results (1.05 = 5% overhead)

src/NetworkOptimizer.Sqm/ScriptGenerator.cs

@@ -278,8 +278,8 @@ private string GenerateSpeedtestScript(Dictionary<string, string> baseline)
         sb.AppendLine(GetTcUpdateFunction());
         sb.AppendLine();
 
-        // Set probe rate slightly above line rate before speedtest so TC never engages
-        sb.AppendLine("# Set SQM to probe rate (~5% above line rate) before speedtest for truly unshaped measurement");
+        // Set probe rate slightly above max shaping rate before speedtest so TC never engages
+        sb.AppendLine("# Set SQM to probe rate (3% above max shaping rate) before speedtest for unshaped measurement");
         sb.AppendLine("update_all_tc_classes $IFB_DEVICE $SPEEDTEST_PROBE_RATE");
         sb.AppendLine("# Upstream: shape rate if enabled, otherwise just tune performance params");
         sb.AppendLine("if [ \"$SHAPE_UPLOAD\" = \"1\" ]; then");
@@ -572,16 +572,14 @@ private string GeneratePingScript(Dictionary<string, string> baseline)
     /// </summary>
     private string GetTcUpdateFunction()
     {
-        return @"# Burst size tuning history:
-# Config E testing showed 5KB burst eliminates downstream drop_overmemory for bulk
-# flows at gig speeds, but 8KB+ creates bursty HTB send patterns (dump at wire rate,
-# pause for token refill) that increase queue depth variance in fq_codel - shows up
-# as latency jitter under load even though fq_codel is working correctly.
-# Whether higher burst helps depends on ISP policing/shaping - some ISPs police with
-# token bucket and penalize bursts above stock, others are fine with 5KB.
-# Reverted to stock 1500b pending per-connection configurability.
+        return @"# 5KB burst eliminates downstream drop_overmemory for bulk flows at gig speeds.
+# 8KB+ creates bursty HTB send patterns that increase queue depth variance in fq_codel.
 calc_burst() {
-    echo ""1500""
+    local rate_mbps=$1
+    local burst=$((rate_mbps * 5))
+    [ ""$burst"" -lt 1500 ] && burst=1500
+    [ ""$burst"" -gt 5000 ] && burst=5000
+    echo ""$burst""
 }
 
 # Calculate fq_codel memory_limit scaled to rate (prevents drop_overmemory at high speeds)