Changelog

After wiping local data and purging the cloud workspace, stale cursor

state in the local cloud_sync_cursors table caused the sync engine to

re-upload entities that plugins recreate on startup. The server-side

purge cleared S3 batches and API cursors, but the local DuckDB cursor

table was never cleared — so when sync restarted and PushAllEntities()

ran, the engine treated the re-created entities as new changes.

Adds privstack_cloudsync_clear_cursors FFI export that calls the

existing EntityStore::clear_cloud_cursors() method, wires it through

NativeLibrary → ICloudSyncService → CloudSyncService, and calls it

in SeedDataService.PurgeCloudWorkspaceIfConnectedAsync() immediately

after the server-side purge succeeds.

When the user wipes data (Wipe, Reseed, or Reset Data on unlock screen),

the local DuckDB cursors are destroyed but the server-side sync_batches

and sync_cursors remain. On next sync startup the device's cursor resets

to 0, causing getPendingChanges to report the device's own old batches

as pending — triggering an infinite download loop of stale data every

poll cycle.

Fix: call the server's /api/cloud/workspace/purge endpoint before wiping

local data. This clears batches, cursors, snapshots, blobs, and S3

objects server-side so the sync engine starts clean.

• SeedDataService: add PurgeCloudWorkspaceIfConnectedAsync() as step 0

in WipeAllPluginDataAsync (covers both Wipe and Reseed paths)

• SetupWizardViewModel: add PurgeCloudWorkspaceAsync() before runtime

shutdown in ConfirmResetData (covers unlock screen Reset Data)

• Both methods are best-effort — cloud purge failure doesn't block

local wipe

When Cmd+M is pressed without a text input focused, the speech recording

now proceeds normally and routes the transcription to Duncan's AI tray

instead of silently discarding it. This enables voice commands:

• MainWindow.Speech.cs: removed early return when no text control is

focused — always start recording regardless of focus state

• MainWindow.axaml.cs: OnTranscriptionReady routes to

AiTrayVM.SendVoiceMessageAsync() when _speechTargetControl is null,

and auto-opens the AI tray panel

• AiSuggestionTrayViewModel.Chat.cs: added SendVoiceMessageAsync() that

sets ChatInputText and triggers the standard send pipeline (RAG +

intent classification)

Flow: Cmd+M → record → Whisper transcribe → no text field? → Duncan

processes via RAG + intent engine → shows response/action in AI tray.

Migrate remaining #[cfg(test)] inline test modules to tests/ directories.

Crates migrated:

• privstack-crdt: 18 pn_counter tests → tests/pn_counter_tests.rs
• privstack-crypto: 5 recovery tests → tests/recovery_tests.rs
• privstack-ffi: 337 tests → tests/ffi_tests.rs (biggest migration)

FFI crate changes:

• Added "rlib" to Cargo.toml crate-type for integration test linking
• Made DTOs pub: SyncEventDto, DiscoveredPeerInfo, SyncStatus,

CloudFileInfo, LicenseInfo, ActivationInfo, FfiDeviceInfo,

SdkRequest, SdkResponse (with pub fields)

• Made functions pub: init_core, init_with_plugin_host_builder,

license_error_to_ffi, nullable_cstr_to_str, to_c_string

• Made EntityRegistry methods pub: new, register_schema, get_schema,

get_handler, has_schema

• Made HANDLE static pub, plugin_ffi module pub

Remaining #[cfg(test)] in cloud/config.rs and crypto/key.rs are

test-only constructors (CloudConfig::test, KdfParams::test), not

inline test modules.

Total: 360 integration tests, 0 inline test modules remaining.