docs: record Bluetooth hang findings and resume workspace planning

docs/constellation-work.md

@@ -133,14 +133,16 @@ start/stop commands. Niri 26.04 itself has no key-release bindings.
 Noctalia suppresses microphone activity popups/indicators; camera and screen-sharing
 indications remain. Automatic paste and live capture need interactive checks.
 
-Hang investigation: stopping live Vulkan transcription never completed; cancel
-returned the UI to idle but subsequent requests reported pipeline busy. Handy was
-no longer running when inspected; no crash cause is established. The same saved
-1.56-second recording completed in CPU batch mode (135 ms inference, 729 ms load)
-and Vulkan batch mode (156 ms inference, 420 ms load). This checks model inference,
-not live capture/stop/cancel. No new audio was recorded. Switching to CPU is not a
-confirmed fix. After activation, retry live recording and stopping, and inspect
-the service logs if it hangs.
+Hang investigation, parked at the user's request: a debugger capture confirmed
+Handy's recorder waits for another audio chunk before processing its stop command.
+When capture stalls, stopping can therefore hang indefinitely; the transcription
+worker was idle. Bluetooth transport failures accompanied the hangs, and the user
+reported dictation working after disconnecting the Sony WH-1000XM6. The user wants
+its microphone when connected. Keeping the headset in a microphone-enabled HFP
+profile was proposed to avoid repeated profile transitions, but was not applied
+or verified; the reconnect attempt timed out. No source or Bluetooth configuration
+was changed. CPU and Vulkan batch inference both worked; changing inference backend
+is not an established fix. Resume this investigation only if requested.
 Synthetic listener checks cover release during startup, repeat, modifier-first
 release, device disconnect and cancellation, without live input/audio access.
 System activation and an interactive hold/release trial remain user-owned.
@@ -151,8 +153,9 @@ The focus is the broader environment, not further API/database polish. Candidate
 areas remain places/ownership, navigation/interaction, human/agent collaboration,
 and system operation/maintenance; these are discussion areas, not a work queue.
 
-Current choice: keep Herdr and configure its keys. The user likes its structure,
-spaces and overview; keyboard interaction is the issue. Agent evaluation is
+Current focus: resume project/workspace organization after the dictation detour.
+Herdr's keys are configured below; their practical fit still needs user feedback.
+The user likes its structure, spaces and overview. Agent evaluation is
 explicitly deferred. Do not resume replacement-tool comparisons or ask for an
 evaluation task/budget as a prerequisite for this work.