Share Speech Dispatcher transport and restart recovery

README.md

@@ -91,3 +91,38 @@ pinned module versions: after review and live acceptance, release/tag the
 common module, then update `go.mod` and `go.sum` in **both** readers and test
 with `GOWORK=off`. A shared source change does not update already installed
 binaries; they must be rebuilt with the updated dependency.
+
+## Speech Dispatcher
+
+`speechdispatcher` owns the shared SSIP client used by both readers: endpoint
+resolution, optional daemon startup, speaking and independent cancellation
+connections, settings, bounded reply parsing, discovery queries, and read-all
+notifications. It has no service-manager dependency. Each reader supplies its
+client identity and keeps its own admission, interruption, presentation, and
+hardware-synthesizer behavior. Fenrir enables automatic system-socket fallback;
+explicit socket/address choices never fall back to another server.
+
+After a connection fails, the failed utterance is not replayed. The next fresh
+speech request rebuilds both channels and restores settings, priority, and the
+new cancellation target. Failed reconnection remains retryable on subsequent
+requests. Context cancellation bounds connection work; Close and Abort prevent
+reconnection. Fenrir additionally retains its worker's bounded background retry
+and outage-queue discard policy. An already accepted utterance whose reply was
+lost may have spoken before the failure; automatically replaying it is unsafe.
+
+`Speak` normalizes desktop text; `SpeakText` preserves console-prepared whitespace
+and line boundaries. `SpeakCharacter` sends character feedback. Query connections
+are separate from speaking connections, and ordinary discovery never autostarts
+a daemon. Startup can honor the invoking sudo user's credentials without
+changing the reader process's credentials.
+
+Run `go test -race ./speechdispatcher` for protocol and recovery fixtures.
+`CTHULHU_NATIVE_SPEECH_STARTUP=1 go test -race ./speechdispatcher` additionally
+starts and restarts private dummy-output daemons; it does not restart the user's
+Speech Dispatcher. The environment variable retains the migrated test's name.
+
+The shared client is available starting with v0.3.0. Both readers pin this
+version for standalone builds; the sibling workspace remains useful for local
+development. Verify released dependencies with `GOWORK=off`. Existing installed
+binaries do not gain fixes until rebuilt. Live audible-speech, keyboard, and
+braille acceptance is tracked separately from automated protocol tests.