CallVault v2.4.1

Added

• A search line at the top of Recordings, Transcripts and Summaries. On Recordings, type a name and
  the list narrows to that contact, with the names of contacts you have recordings with offered as you
  type; it replaces the Contact filter chip. On Transcripts and Summaries, type a word and only the calls
  in which it was said remain, each showing the words around the match. Words match from their start, so
  the list narrows as you type. The magnifier in the header, and the search sheet it opened, are gone —
  the search is on the page itself.

• An option to start recording an outgoing call only when the other person answers. Off by default,
  under Settings → Recording → Outgoing calls. With it on, an automatically recorded phone call you
  place begins at the pickup instead of at the dial, so the file holds the conversation and not the
  ringing, and a call nobody answers leaves no file. App calls too, when the app shows a call timer on
  its notification — WhatsApp does; an app that shows none is recorded from the start as before. If the
  phone cannot say whether the call was answered, it is recorded from the dial as before — the option
  never costs a call.

• Shizuku recordings get speaker labels (#38). A Shizuku recording is already stereo — you on one
  side, the other person on the other — but CallVault had no way to read that, so transcripts came out
  as unattributed text. It now works out who spoke from the file itself while transcribing it, so the
  labels appear on the transcript exactly as they do in built-in mode. Recordings made before this
  update get them too: open one and choose Transcribe again.

Fixed

• App-call recordings on Android 14 and earlier now carry the contact's name. The name comes from the
  calling app's own call notification, and on Android 14 the way the system describes that notification
  differs from Android 15 and later — CallVault only understood the newer form, so every WhatsApp, Telegram
  or Signal recording on an Android 14 phone was named by time alone. Both forms are read now, and if the
  notification is not there yet at the instant the call starts, CallVault looks again a few seconds in.

• An update in Shizuku mode no longer costs you the next call. Installing an update left the old
  recorder running with a path to an app file that no longer existed. It kept answering, so nothing
  looked wrong — and the next call recorded nothing, with no error, until the app was next opened. The
  old recorder is now properly retired as part of the update, and a fresh one is ready before any call
  arrives. As a backstop, every recorder CallVault talks to is asked which app file it was started from,
  and one started from an app that has since been replaced is retired on the spot. If a call is being
  recorded at the moment an update lands, that recording is left alone. Present in every earlier version
  with Shizuku mode, including 2.4.0.

• App-call recordings now report how their two sides lined up (#41). A user hears the other person
  a beat early in WhatsApp recordings. The two sides of an app call are captured separately and
  combined, and until now nothing measured how well; with debug logs on, every app call now logs the
  offset between the two sides as it goes, everything that can shift one side (a silenced or re-taken
  microphone, a dropped or late chunk), and where the sound was playing — earpiece, speaker or
  Bluetooth. Diagnostics only; the recording itself is unchanged. If you hear this, turn on debug logs,
  make one call, and share the log from Settings → Debug.

• A transcript line shared by both speakers is cut where the speaker changes. Whisper decides where a
  line ends by the sentence, not by who is talking, so when one person answered the other inside a single
  line the transcript could show neither name for it, or one long unbroken paragraph. A line is now split
  at a pause where the other person takes over, and the pieces of one person's turn are shown as one row
  each — one row per turn — instead of one row per sentence.

• A transcript that came out in the wrong alphabet is decoded again. With the language pinned to
  Hebrew (or Arabic, Russian, Chinese) a short call could still come back in English on some phones —
  the language setting is a strong hint to the recogniser, not a guarantee. When the text comes back in
  Latin letters under a non-Latin pin, the recording is transcribed again with different decoding
  settings and the first clean result is kept. Recordings longer than three minutes are not retried.

• A transcript line could go unlabelled even though only one person spoke in it. On some phones a
  voice is faintly picked up on the other side's channel as well, and those moments were being read as
  both people talking at once — enough of them and the line was left without a name. A speaker is now
  named when they are the only one who spoke alone in that line.