EchoMuse v2.9.9

On-device wake word scoring (experimental, off by default)

This Echo can now run the wake-word model itself and report what it would
have detected — without acting on it. Nothing about how your Echo behaves
changes unless you switch it on, in Config → Wake word → "Score on device".

It exists to answer whether on-device detection is trustworthy before
anything depends on it. With it on, each turn's row in the Activity tab shows
the Echo's own score beside the controller's, and the activity view reports
how often they agreed and how far apart they were.

Before you enable it:

• It needs files placed on the Echo that are not part of this firmware —
  ONNX Runtime and the wake-word models, about 15MB, in
  /data/local/share/echomuse/oww. They are deliberately not shipped in the
  firmware: that would double both this download and the space each of the
  two firmware slots takes. Until they are there the toggle does nothing,
  and the device log names the file it could not find.

• It costs roughly half a CPU core, permanently, because the wake stream is
  always on. Measured as comfortable on this hardware — the microphone
  pipeline was unaffected across hours of use, including during music
  playback — but turn it on for one Echo at a time and watch the Resources
  panel on the Status tab.

• It needs a controller new enough to record the comparison. An older
  controller simply never switches it on, so there is no bad pairing.

Also in this release

• A second CPU core is kept awake. The Echo has four and parks the ones it
  is not using; the audio pipeline works to a 160ms deadline and does
  better when it is not sharing with everything else.

• The Status tab now shows the Echo's temperature and how many of its CPU
  cores are awake, and warns if the chip ever begins limiting itself
  thermally. It idles at around 33°C, so anything amber there is genuinely
  unusual.

Updating is safe and needs nothing from you. If an update ever goes wrong,
the Echo rolls back to its previous version on its own.

What's Changed

• api: config POSTs refuse to silently delete settings by @wilbowes in #26
• fix: config guard stole @auth.require_admin from _post_global_config by @wilbowes in #27
• ring: legible playback meter, honest turn-state, tunable curve by @wilbowes in #29
• config: per-section fleet/device scoping, plus a last_seen fix by @wilbowes in #30
• instrumentation + media player fixes: RTT, resume seek, feed buffer by @wilbowes in #31
• Save and play back the mic audio of recent turns by @wilbowes in #42
• Tap the utterance recording below NS, not above it by @wilbowes in #43
• docs: record why delay-and-sum is marginal on a 72mm array by @wilbowes in #44
• wakeword: streaming feature pipeline, validated against Python by @wilbowes in #45
• wakeword: run the models on the device, via a dlopen'd ONNX Runtime by @wilbowes in #46

Full Changelog: v2.9.7...v2.9.9