Set _completion_triggered on unseen-RUNNING terminal states (#790)

  When a printer enters FINISH/FAILED without Bambuddy ever seeing
  RUNNING (e.g. after a restart), mark _completion_triggered = True
  so the flag is clean for the next print cycle.

CHANGELOG.md

@@ -29,7 +29,7 @@ All notable changes to Bambuddy will be documented in this file.
 - **SpoolBuddy Update UI Stale After Restart** — After a SpoolBuddy update, the UI permanently showed the old version and "update available" because: (1) the SSH update set status to `"complete"` after the daemon had already re-registered, overwriting the cleared state; (2) the kiosk restart navigated away from the updates page; (3) query cache served stale data. Fixed by letting daemon re-registration clear all update status, removing the kiosk restart in favor of a frontend-driven `window.location.reload()` triggered via WebSocket when the daemon comes back online, and adding proper loading states to Check/Force Update buttons.
 - **Virtual Printer Proxy A1 Printing Fails** ([#757](https://github.com/maziggy/bambuddy/issues/757)) — BambuStudio could not send prints to A1 (and potentially P1S) virtual printers in proxy mode. The slicer connects to undocumented proprietary ports 2024-2026 on these models, which the proxy was not forwarding, causing BambuStudio to show an access code dialog instead of printing. Added transparent TCP pass-through proxying for ports 2024-2026. These ports are silently ignored on models that don't use them (X1C, H2C, P2S). Reported by @Utility9298.
 - **Spool Assignment on Empty AMS Slots** ([#784](https://github.com/maziggy/bambuddy/issues/784)) — Empty AMS slots (no physical spool detected) showed "Assign Spool" and "Configure" buttons in the hover popup. Assigning a spool to an empty slot created a stuck state because no "Unassign" button is available for empty slots. Truly empty slots now hide both buttons, while slots with a spool inserted but filament not loaded still show configure/assign. Also fixed stale AMS slot data on H2D and other printers that only send `{id, state}` in incremental MQTT updates — filament load/unload transitions now update in real-time without requiring a reconnect. Reported by @RosdasHH.
-- **Log Flood: "State is FINISH but completion NOT triggered"** ([#790](https://github.com/maziggy/bambuddy/issues/790)) — A diagnostic log message introduced in 0.2.2.1 fired on every MQTT update while a printer sat in FINISH or FAILED state, flooding logs with thousands of lines per minute in printer farms. Fixed by only logging once on the initial state transition. Reported by @user.
+- **Log Flood: "State is FINISH but completion NOT triggered"** ([#790](https://github.com/maziggy/bambuddy/issues/790)) — A diagnostic log message introduced in 0.2.2.1 fired on every MQTT update while a printer sat in FINISH or FAILED state, flooding logs with thousands of lines per minute in printer farms. Fixed by only logging once on the initial state transition, and marking `_completion_triggered = True` when a terminal state is first seen without a prior RUNNING state so the flag is clean for the next print cycle. Reported by @user.
 - **H2D External Spool Print Fails With "Failed to get AMS mapping table"** ([#797](https://github.com/maziggy/bambuddy/issues/797)) — Printing from an external spool on H2D (and H2D Pro) through Bambuddy failed with `0700_8012 "Failed to get AMS mapping table"`, while the same print worked fine from BambuStudio. Bambuddy was passing raw virtual tray IDs (254/255) in the flat `ams_mapping` array, but BambuStudio converts these to -1 and relies on `ams_mapping2` for external spool routing. The H2D firmware rejects raw 254/255 in the flat array. Also fixed the `ams_mapping2` format for external trays — each virtual tray is its own AMS unit with `slot_id: 0`, not a shared unit differentiated by slot. Reported by @Lukas-ESG.
 - **SpoolBuddy Scale First Reading Always Wrong** — The NAU7802 ADC always returns a stale max-scale value (`0x7FFFFF`) on its first conversion after power-up, which polluted the moving average and made the initial weight report wildly inaccurate. Fixed by flushing the first reading during `init()` so all subsequent reads return valid data. Also extracted both hardware drivers out of diagnostic scripts into proper modules — the NAU7802 scale driver from `scripts/scale_diag.py` into `daemon/nau7802.py`, and the PN5180 NFC driver from `scripts/read_tag.py` into `daemon/pn5180.py`. The production daemon was importing driver classes from test scripts since the original SpoolBuddy commit. Removed the now-unnecessary `sys.path` hack from `main.py`.
 - **ffmpeg Process Leak Causing Memory Growth** ([#776](https://github.com/maziggy/bambuddy/issues/776)) — Camera stream ffmpeg processes accumulated over time, consuming several GB of RAM. When a user closed the camera viewer, the frontend sent a stop signal that killed the ffmpeg process, but the backend stream generator interpreted the dead process as a dropped connection and respawned ffmpeg — up to 30 reconnection attempts per stream. The orphan cleanup couldn't catch these because they were tracked as "active". Fixed by signaling the generator's disconnect event from the stop endpoint before killing the process, checking for stream removal before reconnecting, and tracking frame timestamps per-stream instead of per-printer so stale detection works correctly when multiple streams exist. Reported by @ChrisTheDBA, confirmed by @peter-k-de.

backend/app/services/bambu_mqtt.py

@@ -2466,6 +2466,8 @@ class BambuMQTTClient:
                 f"prev={self._previous_gcode_state}, was_running={self._was_running}, "
                 f"already_triggered={self._completion_triggered}, has_callback={bool(self.on_print_complete)}"
             )
+            # Mark as triggered so state is clean for the next print cycle
+            self._completion_triggered = True
 
         if should_trigger_completion:
             if self.state.state == "FINISH":