Add server-side Piper read-aloud with voice picker

Reader read-aloud now streams neural WAV audio from a new POST /api/tts
endpoint that shells out to Piper, instead of the browser's Web Speech
voice. Each paragraph is synthesized on demand with the next one
prefetched during playback, keeping the existing highlight/scroll sync.

Voices are configured under [web.tts] (piper binary + voices_dir + a
labelled voice list) and exposed to the client as window.PETE_TTS; the
reader gets a Voice selector in the Aa menu, persisted per-device. Still
a signed-in-only perk and gated on auth.

config.example.toml

@@ -77,6 +77,25 @@ interval_minutes = 360
 # item doesn't ping everyone.
 min_stories = 3
 
+# Server-side neural read-aloud (Piper, https://github.com/rhasspy/piper).
+# When enabled, the reader's "Listen" button streams real Piper voices instead
+# of the browser's robotic Web Speech voice. Signed-in only, so it needs
+# [web.auth] on too. Install the piper binary and one or more voice models
+# (<id>.onnx + <id>.onnx.json) into voices_dir first.
+[web.tts]
+enabled = false
+piper_bin = "/opt/piper/piper"      # path to the piper executable
+voices_dir = "/opt/piper/voices"    # dir holding <id>.onnx (+ .onnx.json) models
+default = "en_US-amy-medium"         # voice id selected until the reader picks another
+# List the voices to offer, in menu order. Omit the whole [[web.tts.voices]]
+# list to auto-discover every *.onnx in voices_dir (labelled by filename).
+[[web.tts.voices]]
+id = "en_US-amy-medium"
+label = "Amy (US, female)"
+[[web.tts.voices]]
+id = "en_US-ryan-high"
+label = "Ryan (US, male, HQ)"
+
 # Every enabled source MUST set direct_route to a key from [matrix.channels] above.
 # There is no automatic classification — Pete posts each story to its configured channel.
 # Optional: language = "en" drops feed items whose per-item <language> tag is