fix(cw): a drifting tone wiped the record instead of filling it

The record could go from a screenful of text to less, and then to nothing at all, while
decoding was still running. The cause is dropBufferedAudio(), which runs on every change
of shift - and the shift tracks the detected tone, which drifts across a pass.

The live window is the only route into the archive: audio gets there by being pushed out
by newer audio. Clearing the window therefore did not merely discard 20 s, it reset the
progress towards ever archiving anything. Modelled at 18 WPM, a drift every 20 s meant
five minutes of listening archived not one character, however long the operator waited -
the window was always wiped before the first sample could be evicted. With the record
still concatenating the live decode at the time, each wipe also cut the visible
transcript short, which is the text going backwards and then never accumulating.

Retiring the window instead of dropping it fixes both. Those samples cannot stay - one
spectrogram over two shift amounts smears the tone - but they were shifted consistently
and they are complete, so they decode fine on their own. They are queued and archived by
the normal path, keeping dropBufferedAudio() non-suspending: both callers sit on the
synchronous capture path, and decoding there would put an inference inside the tone
scan. Modelled over 300 s, a drift every 5 s goes from 0 characters archived to 449.

An earlier attempt at this - keeping archiveSize instead of zeroing it - was wrong and
the probe rejected it: with the window wiped before it ever overflowed, that buffer was
empty, so preserving it preserved nothing.

The queue is synchronised (written from capture, drained from capture and flush) and
capped at four windows, dropping the oldest when full: a tone drifting on every
detection scan would otherwise queue faster than the decoder can drain.

Also from an audit of the previous two commits:

- The copy button was gated on the record, which is empty for the opening half-minute
  while the first batch accumulates. That greyed out the one control that rescues the
  text over exactly the short exchange most likely to be lost. It now copies the live
  window too, and the empty state says which surface updates sooner.
- cw_copy, cw_copied, cw_record_label and cw_record_empty were missing from values-id,
  values-in and values-tr, which carry the full UI strings, so those users saw English.
- Dropped a KDoc block left dangling by 8445c170, which had removed the constant it
  documented.

core/data/src/main/java/com/rtbishop/look4sat/core/data/cw/CwDeepDecoder.kt

@@ -166,6 +166,23 @@ class CwDeepDecoder(
     private val archiveBuffer = FloatArray(CwDeepBuffer.DEFAULT_MAX_SECONDS.toInt() * CwDeepSpectrogram.SAMPLE_RATE)
     private var archiveSize = 0
 
+    /**
+     * Window contents retired by a change of shift, waiting to be archived.
+     *
+     * The live window is the only route into the archive - audio gets there by being pushed
+     * out - so clearing the window used to mean its audio was never decoded at all. When the
+     * shift changed more often than the window took to fill, that was every sample: modelled
+     * at 18 WPM with a drift every 20 s, five minutes of listening archived nothing whatever.
+     * Synchronised on itself: written from the capture path and drained from both there and
+     * [flush], which run on different coroutines. Capped, because a signal drifting on every
+     * detection scan would otherwise queue windows faster than they can be decoded and grow
+     * without bound; past the cap the oldest goes, since newer audio is what is being read.
+     */
+    private val retiredAudio = ArrayDeque<FloatArray>()
+
+    /** Windows held awaiting archival before the oldest is dropped. */
+    private val retiredAudioLimit = 4
+
     /** Held while inference runs so slow devices skip work instead of queuing it. */
     private val inferenceLock = Mutex()
 
@@ -312,6 +329,21 @@ class CwDeepDecoder(
             bandLimited, sampleRate, CwDeepSpectrogram.SAMPLE_RATE
         )
         val prepared = applyToneShift(resampled)
+
+        // Anything applyToneShift just retired from the window is older than what follows, so
+        // it is archived before the new audio is buffered - otherwise the record comes out
+        // with its text transposed. Archived whole rather than accumulated: it is already a
+        // full window's worth, and holding it back would only expose it to the next retirement.
+        val retired = drainRetiredAudio()
+        for (batch in retired) {
+            try {
+                archiveDecode(batch)
+            } catch (t: Throwable) {
+                if (t is CancellationException) throw t
+                Log.e(TAG, "retired audio decode failed", t)
+            }
+        }
+
         val shouldRedecode = buffer.append(prepared)
 
         // Archive audio that scrolled out of the live window. It is decoded once
@@ -438,9 +470,49 @@ class CwDeepDecoder(
      * cannot be un-shifted, so they are dropped rather than decoded against the new
      * shift. Text already committed to [historyText] stays: it was correct when decoded.
      */
+    /**
+     * Drop audio that was shifted by a setting no longer in force - but keep what can be kept.
+     *
+     * The live window has to go: it holds samples moved by two different amounts, and one
+     * spectrogram over both smears the tone. The pending archive batch does not. Those samples
+     * already left the window, they were shifted consistently, and they are complete, so
+     * discarding them threw away decodable audio for no reason. They are left in place here to
+     * be archived by the normal path, which keeps this function non-suspending: its two
+     * callers sit on the synchronous capture path, and making them suspend to run an inference
+     * here would put a decode inside the tone-detection scan.
+     *
+     * It mattered because this runs on every change of shift, which tracks the detected tone,
+     * which drifts across a pass. Modelled at 18 WPM, a drift every 20 s left the record
+     * permanently empty however long the operator listened: the window was wiped before any
+     * batch could complete, so nothing was ever committed. With the record still concatenating
+     * the live decode at the time, each wipe visibly cut the transcript short as well - text
+     * going backwards, then never accumulating at all.
+     */
+    /** Take every retired window, oldest first, leaving the queue empty. */
+    private fun drainRetiredAudio(): List<FloatArray> = synchronized(retiredAudio) {
+        if (retiredAudio.isEmpty()) {
+            emptyList()
+        } else {
+            retiredAudio.toList().also { retiredAudio.clear() }
+        }
+    }
+
     private fun dropBufferedAudio() {
+        // Retired, not discarded. The samples cannot stay in the window - mixing two shifts
+        // in one spectrogram smears the tone - but they are internally consistent and
+        // complete, so they decode fine on their own. Handed to the archive path rather than
+        // decoded here, because both callers sit on the synchronous capture path.
+        val retiring = buffer.snapshot()
+        if (retiring.isNotEmpty()) {
+            synchronized(retiredAudio) {
+                while (retiredAudio.size >= retiredAudioLimit) {
+                    Log.w(TAG, "retired audio queue full, dropping the oldest window")
+                    retiredAudio.removeFirst()
+                }
+                retiredAudio.addLast(retiring)
+            }
+        }
         buffer.reset()
-        archiveSize = 0
         // Committed text stays: it was correct for audio that really was archived.
         _historyText.value = committedText
     }
@@ -539,6 +611,12 @@ class CwDeepDecoder(
      * has to be archived first or the record comes out with its text transposed.
      */
     override suspend fun flush() = archiveLock.withLock {
+        // Oldest first, all the way down: retired window contents, then the batch accumulating
+        // towards the threshold, then what is still live.
+        for (batch in drainRetiredAudio()) {
+            runCatching { archiveDecodeLocked(batch) }
+                .onFailure { if (it is CancellationException) throw it }
+        }
         if (archiveSize > 0) {
             val pending = archiveBuffer.copyOf(archiveSize)
             archiveSize = 0
@@ -660,6 +738,7 @@ class CwDeepDecoder(
     override fun reset() {
         antiAlias?.reset()
         buffer.reset()
+        synchronized(retiredAudio) { retiredAudio.clear() }
         _decodedText.value = ""
         _historyText.value = ""
         committedText = ""