Fix Piper memory pressure on iOS

docs/adr/0003-limit-piper-wasm-to-one-thread.md

@@ -0,0 +1,13 @@
+# ADR 0003: Limit Piper WASM to one thread and reset failed workers
+
+## Decision
+
+The Piper worker reports one available processor to Piper and configures ONNX Runtime for one thread. If a synthesis request fails, the browser client terminates that worker; the next player retry creates a fresh worker and ONNX module.
+
+## Context
+
+Piper TTS Web 1.0.3 sets `env.wasm.numThreads` from `navigator.hardwareConcurrency` during `TtsSession.init()`, after the application had configured ONNX Runtime. On iPad this could undo the single-thread setting and request a larger WASM worker pool. The library also retains a singleton session if ONNX initialization fails, so retrying in the same worker can reuse a failed initialization.
+
+## Consequences
+
+The one-thread limit reduces parallel inference but lowers memory pressure on mobile Safari. Failed initialization no longer poisons future attempts in the same worker. This cannot guarantee that every device has enough memory for the 63 MB medium voice model; the CI release and Kubernetes checks validate packaging and deployment, while actual iPad memory behavior still requires device verification.