Guard local rewrites and document F-Droid limits

CHANGELOG.md

@@ -5,6 +5,29 @@ All notable changes to Rescripto are documented in this file.
 The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project uses [Semantic Versioning](https://semver.org/).
 
+## [1.3.2] - 2026-10-27
+
+### Changed
+
+- Rewrite instructions now explicitly preserve the draft's language and time
+  format, including mixed-language passages. A clear translation or number
+  change is rejected before it reaches history. On-device rewrites can retry
+  a rejected draft by language passage; the original text remains available
+  if the model still cannot preserve it. Small models can still make mistakes;
+  review the result before using it.
+- Added a hint beside the intensity controls explaining that Full rewrite can
+  produce stronger tone changes, especially with small models. Simple mode also
+  explains where to find this setting.
+- Clarified the store description that preservation is an instruction to the
+  model, not a guarantee, and that results should be checked before use.
+- Clarified that on-device rewrites can take minutes on some phones and that
+  model speed varies by device.
+- Added an on-screen reminder to compare AI results with the original before
+  sending. A missed-meeting safeguard also rejects an English rewrite that
+  silently drops the speaker's inability to attend.
+- Silence-only local dictation no longer adds a `[ Silence ]` marker to the
+  draft.
+
 ## [1.3.1] - 2026-09-16
 
 ### Changed

docs/FDROID.md

@@ -19,7 +19,8 @@ draft build recipe, and the steps that still need a person.
 
 ## Anti-features to declare
 
-- **NonFreeNet**: optional. In Cloud or Hybrid mode, text and recordings can
+- **NonFreeNet**: local use requires downloading an AI model from Hugging Face.
+  In optional Cloud or Hybrid mode, text and recordings can
   go to commercial AI services (OpenAI, Anthropic, Google, Groq, xAI and
   others). All of it is off until the user turns it on and adds their own
   API key.
@@ -41,8 +42,8 @@ to time.
 ```yaml
 AntiFeatures:
   NonFreeNet:
-    en-US: Can send text to commercial cloud AI services when you turn on Cloud or
-      Hybrid mode.
+    en-US: Requires downloading an AI model from Hugging Face for local use and can
+      send text to commercial cloud AI services in Cloud or Hybrid mode.
 Categories:
   - Writing
 License: Apache-2.0

docs/FDROID_DEVICE_RETEST.md

@@ -0,0 +1,54 @@
+# F-Droid device retest
+
+Related review: https://gitlab.com/fdroid/fdroiddata/-/merge_requests/49009#note_4067917436
+
+The isolated QA build was tested on a Redmi 12 (Android 15), leaving the
+installed production 1.3.0 app and its data intact. See
+[`testing/mobile/reports/2026-10-25_fdroid_language_retest.md`](../testing/mobile/reports/2026-10-25_fdroid_language_retest.md)
+for the inputs, complete outputs, model versions, timings, and screenshots.
+
+## Current result
+
+- Gemma 3 1B and Qwen 2.5 1.5B kept the F-Droid reviewer's German sample in
+  German, including `13 uhr` and the train cancellation. A German/English
+  sample retained its language switches. French, Spanish, and one Chinese
+  sample also retained their source languages. This sample set cannot establish
+  reliable behavior for every language.
+- The exact English meeting draft was checked at Casual / Full and
+  Professional / Full. Gemma preserved the observed facts and saved a modest
+  rewrite in about 35 seconds. Qwen omitted or changed facts, and the app
+  rejected both of its generated rewrites. Professional / Full also produced
+  two immediate empty responses until the QA app was restarted.
+- Llama 3.2 1B was downloaded and verified as an English-only alternative.
+  Professional / Full preserved the facts but was a slight, informal edit;
+  Casual / Full omitted the reports and was rejected. It did not improve the
+  English tone and fidelity tradeoff in these runs.
+- Japanese exposed a semantic failure: Gemma changed “I cannot attend the
+  meeting” to “the meeting cannot be held” while retaining Japanese script,
+  date, time, and number. The current guard cannot detect that meaning change.
+  Do not claim that all languages or tones are accurate.
+- A silent dictation recording previously appended Whisper's `[ Silence ]`
+  marker. The QA build now leaves the draft untouched for a silence-only
+  transcript. A spoken retry captured and inserted both sentences the user
+  said, but misheard “Thursday” as “Lutersday.” The draft remained editable.
+  History, sharing, Quick Settings, custom tones and audiences, privacy
+  controls, and a one-step workflow were also exercised.
+
+At `fc7dbab`, 476 Windows Flutter tests and analysis pass, and an Android
+arm64 release-style APK builds. Encrypted backup export/restore, WebDAV
+conflict handling, and a mock OpenAI-compatible cloud stream were exercised
+on the phone against local endpoints. Real cloud and Nextcloud accounts were
+not tested. Android registers Rescripto for selected-text processing, but
+this Redmi's selection menu did not expose it, so insertion back into another
+app remains unverified. No 1.3.2 release or F-Droid recipe update has been
+published.
+
+## Release gate
+
+Local model output is best effort; 100% semantic or tone accuracy is not a
+release requirement. Keep the English, Japanese, speech, and selected-text
+limitations visible to reviewers and users, and do not claim universal
+language support. Before updating the F-Droid recipe, obtain a tested
+immutable release and verify reproducibility against its signed APK. Capture
+actual phone screenshots for the listing if the optional screenshot request
+is addressed.

fastlane/metadata/android/en-US/changelogs/212.txt

@@ -0,0 +1,4 @@
+- Local rewrites better preserve the draft's language and reject clear translation or number changes
+- A result reminder asks you to compare AI text with the original before sending
+- Full rewrite can show stronger tone changes; local model quality and speed vary by device
+- Silence-only dictation no longer inserts a silence marker

fastlane/metadata/android/en-US/full_description.txt

@@ -1,4 +1,4 @@
-<p>Paste a rough message, pick a tone, and Rescripto rewrites it into something you'd actually send. Your facts, names, numbers and dates stay as you wrote them, and the result doesn't get padded with greetings, filler or hype.</p>
+<p>Paste a rough message, pick a tone, and Rescripto rewrites it into something you'd actually send. It is instructed to preserve your language, facts, names, numbers and dates, without adding greetings, filler or hype. Models can still make mistakes, especially smaller on-device models, so check the result before using it.</p>
 
 <p>Rewriting runs on your phone. Download a model once (Gemma 3 1B, Llama 3.2 1B or 3B, or Qwen 2.5 1.5B) and it works offline, with no account.</p>
 
@@ -8,7 +8,7 @@
 <li>In Pro mode, set how much to change, shorter or longer, who it's for, extra instructions, and up to three versions.</li>
 <li>Dictate with whisper.cpp running on the phone.</li>
 <li>Chain rewrite steps into workflows.</li>
-<li>Rewrite selected text from any app's text menu, share text in, or start from a Quick Settings tile.</li>
+<li>Rewrite selected text from Android's text menu when available, share text in, or start from a Quick Settings tile.</li>
 <li>Search and reuse past rewrites. History stays on the phone.</li>
 <li>Export encrypted backups, keep automatic weekly backups, or sync an encrypted copy to your own WebDAV server.</li>
 </ul>

lib/core/constants.dart

@@ -19,7 +19,7 @@ abstract final class AppConstants {
   static const String issuesUrl =
       'https://github.com/Gr33nOps/Rescripto/issues';
   // Keep in step with `version:` in pubspec.yaml.
-  static const String versionName = '1.3.1';
+  static const String versionName = '1.3.2';
 
   // Storage keys (shared_preferences).
   static const String keySelectedModel = 'selected_model_id';

lib/engine/engine_error_messages.dart

@@ -33,6 +33,9 @@ String describeEngineError(EngineException error) => switch (error) {
     'The model wouldn’t rewrite this text, even on a second try. Small '
         'on-device models sometimes do that with ordinary writing. A '
         'different model in the Models tab usually helps.',
+  RewriteFidelityException() =>
+    'The model changed the draft’s language or numbers, so its result was '
+        'discarded. Your original text is still here. Try another model.',
   OutOfMemoryException() =>
     'There isn’t enough memory for this model. Try a smaller model, or a '
         'smaller context size in Settings > Performance.',

lib/engine/engine_exception.dart

@@ -72,6 +72,11 @@ final class ModelRefusedException extends EngineException {
   const ModelRefusedException();
 }
 
+/// A local model translated the draft or changed its numeric details.
+final class RewriteFidelityException extends EngineException {
+  const RewriteFidelityException();
+}
+
 /// The request was cancelled before it produced a result.
 final class GenerationCancelledException extends EngineException {
   const GenerationCancelledException();

lib/engine/workflow_runner.dart

@@ -5,6 +5,7 @@ import '../models/rewrite_request.dart';
 import '../models/workflow_definition.dart';
 import '../models/workflow_step.dart';
 import '../services/config_store.dart';
+import '../services/language_fidelity.dart';
 import '../services/prompt_builder.dart';
 import '../services/refusal_detector.dart';
 import '../services/storage_service.dart';
@@ -131,7 +132,9 @@ class WorkflowRunner extends ChangeNotifier {
     await engine.prepare(step.target);
 
     final tone = _configStore.toneById(step.toneId);
-    final audienceLabels = step.audience.map((id) => _configStore.audienceById(id).label).toList();
+    final audienceLabels = step.audience
+        .map((id) => _configStore.audienceById(id).label)
+        .toList();
     final request = RewriteRequest(
       sourceText: inputText,
       toneId: step.toneId,
@@ -149,18 +152,32 @@ class WorkflowRunner extends ChangeNotifier {
       maxOutputTokens: tone.maxOutputTokens,
       stopSequences: tone.stopSequences,
     );
+    final languageProfile = LanguageFidelity.inspect(inputText);
 
     Future<List<String>> attempt({
       bool strictRetry = false,
       bool unchangedRetry = false,
     }) async {
-      final prompt = PromptBuilder.build(
-        request,
-        tone: tone,
-        audienceLabels: audienceLabels,
-        strictRetry: strictRetry,
-        unchangedRetry: unchangedRetry,
-      );
+      final useCompactEnglish =
+          step.target.engineId == 'local.llama' &&
+          languageProfile.languages.length == 1 &&
+          languageProfile.languages.single == 'en' &&
+          !strictRetry &&
+          !unchangedRetry;
+      final prompt = useCompactEnglish
+          ? PromptBuilder.buildEnglish(
+              request,
+              tone: tone,
+              audienceLabels: audienceLabels,
+            )
+          : PromptBuilder.build(
+              request,
+              tone: tone,
+              audienceLabels: audienceLabels,
+              sourceLanguageCues: languageProfile.promptCues,
+              strictRetry: strictRetry,
+              unchangedRetry: unchangedRetry,
+            );
       final handle = engine.start(
         EngineRequest(target: step.target, prompt: prompt, options: options),
       );
@@ -220,6 +237,9 @@ class WorkflowRunner extends ChangeNotifier {
       }
     }
 
+    if (languageProfile.rejects(variants.first)) {
+      throw const RewriteFidelityException();
+    }
     return variants.first;
   }
 

lib/screens/authoring/tone_editor_screen.dart

@@ -499,10 +499,12 @@ class _GenerationSlider extends StatelessWidget {
             ),
           ),
           SizedBox(
-            width: 44,
+            width: 72,
             child: Text(
               display,
               textAlign: TextAlign.end,
+              maxLines: 1,
+              softWrap: false,
               style: TextStyle(color: color),
             ),
           ),

lib/screens/models_screen.dart

@@ -296,9 +296,9 @@ class _ModelTile extends StatelessWidget {
   }
 
   String _descriptor(AiModel model) {
-    if (model.sizeMb <= 900) return 'Fast · works well on most phones';
+    if (model.sizeMb <= 900) return 'Small model · speed varies by phone';
     if (model.sizeMb >= 2500) return 'Highest quality · needs more memory';
-    return 'Balanced quality and speed';
+    return 'Larger model · speed varies by phone';
   }
 }
 

lib/screens/rewrite_screen.dart

@@ -136,6 +136,14 @@ class _RewriteScreenState extends State<RewriteScreen> {
               selectedId: controller.toneId,
               onChanged: controller.setTone,
             ),
+            if (!isPro) ...[
+              const SizedBox(height: 6),
+              Text(
+                'For a stronger change in tone, switch to Pro mode and try '
+                'Full rewrite. Small models may make subtler changes.',
+                style: Theme.of(context).textTheme.bodySmall,
+              ),
+            ],
             if (isPro) ...[
               const SizedBox(height: 20),
               RewriteControls(
@@ -621,6 +629,17 @@ class _StreamingPanel extends StatelessWidget {
                 ).textTheme.bodySmall?.copyWith(color: scheme.onSurfaceVariant),
               ),
             ],
+            if (!cancelling &&
+                stage == EngineStage.streaming &&
+                capabilities.needsLocalInstall) ...[
+              const SizedBox(height: 8),
+              Text(
+                'On some phones, an on-device rewrite can take a few minutes.',
+                style: Theme.of(
+                  context,
+                ).textTheme.bodySmall?.copyWith(color: scheme.onSurfaceVariant),
+              ),
+            ],
             if (text.isNotEmpty) ...[
               const SizedBox(height: 12),
               Text(