fix: recordings buried under hiss — the cheap resampler was folding it in

Reported precisely: on the air the voice is well clear of the noise; in the
recording the noise is louder than the voice. Same audio, so the fault is in how
it is captured, not in what is captured.

AUDCLNT_STREAMFLAGS_SRC_DEFAULT_QUALITY was set alongside AUTOCONVERTPCM. The
name reads like a sensible default and is not: it selects the CHEAP sample-rate
converter, for cases where quality does not matter. Taking a rig's 48 kHz stream
down to our 16 kHz, it folds everything above 8 kHz back into the audio band —
and receiver hiss is mostly high frequencies. The voice keeps its level, the
noise arrives twice.

Dropped, so Windows uses its normal converter, which filters before decimating.
This affects every capture and render path in the app, the voice keyer included.

changelog.json

@@ -9,7 +9,8 @@
       "The play button becomes a stop button while the recording is going out: it cuts the transmission and releases the PTT, and you can send it again straight away.",
       "The \"From radio\" and mic levels take effect as you drag the slider, and the RX level now applies to what you HEAR through OpsLog, not only to recordings.",
       "The padlocks on frequency, band, mode and times no longer flicker under the pointer and refuse the click.",
-      "The callsign field is wider, taking the room from the RST pair."
+      "The callsign field is wider, taking the room from the RST pair.",
+      "QSO recordings are no longer swamped by hiss: the audio was resampled with the cheap converter, folding everything above 8 kHz back onto the voice."
     ],
     "fr": [
       "Les filtres du cluster gagnent NEW POTA et NEW COUNTY, à côté des pastilles de statut existantes.",
@@ -18,7 +19,8 @@
       "Le bouton de lecture devient un bouton d'arrêt pendant l'émission de l'enregistrement : il coupe l'émission et relâche le PTT, et vous pouvez le renvoyer aussitôt.",
       "Les niveaux « From radio » et micro agissent pendant que vous déplacez le curseur, et le niveau RX s'applique désormais à ce que vous ENTENDEZ dans OpsLog, pas seulement aux enregistrements.",
       "Les cadenas de fréquence, bande, mode et heures ne scintillent plus sous le curseur et acceptent le clic.",
-      "Le champ indicatif est plus large, la place venant de la paire RST."
+      "Le champ indicatif est plus large, la place venant de la paire RST.",
+      "Les enregistrements de QSO ne sont plus noyés par le souffle : l'audio était rééchantillonné avec le convertisseur économique, qui repliait tout ce qui dépasse 8 kHz sur la voix."
     ]
   },
   {

internal/audio/engine.go

@@ -80,7 +80,17 @@ func pcmFormat() *wca.WAVEFORMATEX {
 	}
 }
 
-const autoConvert = wca.AUDCLNT_STREAMFLAGS_AUTOCONVERTPCM | wca.AUDCLNT_STREAMFLAGS_SRC_DEFAULT_QUALITY
+// autoConvert lets WASAPI resample between the device format and ours.
+//
+// SRC_DEFAULT_QUALITY used to be set alongside it, and that name is misleading:
+// it selects the CHEAP converter, meant for cases where quality does not matter.
+// Going from a rig 48 kHz stream down to our 16 kHz, it folds everything above
+// 8 kHz back into the audio band — and a receiver hiss is mostly high
+// frequencies. The result was a recording where the noise sat ON TOP of the
+// voice, while the same audio heard live had the voice well clear of it.
+//
+// Without the flag Windows uses its normal converter, which filters first.
+const autoConvert = wca.AUDCLNT_STREAMFLAGS_AUTOCONVERTPCM
 
 // recordPCM captures from a device into 16 kHz mono 16-bit PCM bytes until the
 // stop channel is closed.