docs(meditate): ADR 0038, D15, dreams spoken, as the developer asked

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>

docs/adr/0038-halkyon-meditate.md

@@ -170,6 +170,39 @@ names and pictures are its own). Everything below is Halkyon's own.
   - **Profile's Practice** shows, once the journal holds a dream, a small *Dream
     journal* card: the dreams of this month and the lucid ones of all.
   - Not done: reminders to write (D12), dreams linked to sittings, figures over time.
+- **D15 Dreams spoken** (3 Nov 2026, the developer's request): speech-to-text in the
+  journal's editor, as Learn's assistant field takes questions spoken (Halkyon Learn's
+  ADR 0021, D67: the device's own speech recognition, `speech_to_text`). The agent's
+  recommendations, to be confirmed:
+  - **Shared, not copied**: Learn's `SpeechInput` and `DeviceSpeechInput` move into
+    halkyon_app with their fake, Learn importing them from there unchanged; beside them
+    `Dictation`, generic, which writes what is heard where the cursor stands (in place
+    of a selection, set apart by a space on either side where the text has none) and
+    shows the words so far. Learn's button stays Learn's: it sends, and replaces from
+    where it started. Todo and Train link the plugin now but declare no microphone, so
+    nothing there listens.
+  - **In the editor**: a microphone under the text, where the device recognises speech
+    (none on Linux, as Learn's); listening, it turns into Stop with "Listening…" and the
+    words heard beside it. It listens in the app's language, else the device's own.
+  - **A dream is longer than a question: it restarts by itself** while the button stays
+    on, rather than offering to continue: half awake, a tap after every pause would
+    interrupt the telling, and the button says throughout that it listens. Each listen
+    may pause 30 s and last 10 minutes (Apple's recogniser stops sooner, Android's after
+    a few seconds' silence whatever is asked), and after three listens running hear
+    nothing it stops by itself, so a phone left on the table doesn't listen on.
+  - **Privacy**: on Android Meditate asks for recognition on the device
+    (`DeviceSpeechInput(preferOnDevice: true)`: Android 12's on-device recogniser where
+    the language is installed, else the recogniser with *prefer offline*); should that
+    fail before a word is heard, it listens as Android otherwise does from then on, which
+    may send the sound to Google or the phone's maker. Apple's and Windows' recognisers
+    choose for themselves (the plugin's on-device flag on Apple fails outright where
+    unsupported, so it isn't asked). The instance never hears it. The privacy page says
+    so where Meditate's server serves it (`privacyPagePoints` in `meditate_base.dart`,
+    with what Meditate keeps sealed), as Learn's does.
+  - **Today's card** offers *Speak it* beside it, opening a new dream already listening.
+  - The platforms declare it as Learn's do: `RECORD_AUDIO` and the recognition service's
+    query on Android, the microphone's and speech recognition's usage texts on iOS and
+    macOS, and macOS's `audio-input` entitlement.
 
 ## Steps
 
@@ -181,7 +214,8 @@ names and pictures are its own). Everything below is Halkyon's own.
    sitting's screen (D6), Explore (D10).
 4. **Today and Profile** (3 Nov): D8, D9.
 5. **The dream journal** (3 Nov): D14, the `dream` table and its migration, the Journal
-   tab, a dream's screen, Today's card, Settings' *Journal*, Profile's figures.
+   tab, a dream's screen, Today's card, Settings' *Journal*, Profile's figures; and
+   dreams spoken (D15).
 6. **Released and deployed** (later, as Todo's step 5): `apps/meditate/release.dart`, a key
    of its own, the F-Droid listing, a snap, `deploy/compose.meditate.yaml` at
    meditate.halkyon.eu (its `.env` values the developer has set already, 2 Nov), its