bambuddy/backend/app/services/virtual_printer/manager.py

"""Virtual Printer Manager - coordinates SSDP, MQTT, and FTP services.

Each virtual printer runs its own independent services (FTP, MQTT, SSDP, Bind)
bound to its dedicated IP address, regardless of mode.
"""

import asyncio
import json
import logging
import time
from collections.abc import Callable
from datetime import datetime, timezone
from pathlib import Path
from typing import TYPE_CHECKING

from backend.app.core.config import settings as app_settings
from backend.app.models.virtual_printer import (
    VP_MODE_ARCHIVE,
    VP_MODE_PROXY,
    VP_MODE_QUEUE,
    normalize_vp_mode,
)
from backend.app.services.bambu_mqtt import resolve_external_spools_in_mapping
from backend.app.services.virtual_printer.bind_server import BindServer
from backend.app.services.virtual_printer.certificate import CertificateService
from backend.app.services.virtual_printer.ftp_server import VirtualPrinterFTPServer, compute_passive_port_slice
from backend.app.services.virtual_printer.mqtt_bridge import MQTTBridge
from backend.app.services.virtual_printer.mqtt_server import SimpleMQTTServer
from backend.app.services.virtual_printer.ssdp_server import SSDPProxy, VirtualPrinterSSDPServer
from backend.app.services.virtual_printer.tcp_proxy import SlicerProxyManager, TCPProxy
from backend.app.utils.printer_models import is_nozzle_rack_model

if TYPE_CHECKING:
    from backend.app.services.printer_manager import PrinterManager

logger = logging.getLogger(__name__)


# Mapping of SSDP model codes to display names
# These are the codes that slicers expect during discovery
# Sources:
#   - https://gist.github.com/Alex-Schaefer/72a9e2491a42da2ef99fb87601955cc3
#   - https://github.com/psychoticbeef/BambuLabOrcaSlicerDiscovery
VIRTUAL_PRINTER_MODELS = {
    # X1 Series
    "BL-P001": "X1C",  # X1 Carbon
    "BL-P002": "X1",  # X1
    "C13": "X1E",  # X1E
    # X2 Series
    "N6": "X2D",  # X2D
    # A2 Series (single-FDM + integrated cutter/plotter)
    "N9": "A2L",  # A2L
    # P Series
    "C11": "P1P",  # P1P
    "C12": "P1S",  # P1S
    "N7": "P2S",  # P2S
    # A1 Series
    "N2S": "A1",  # A1
    "N1": "A1 Mini",  # A1 Mini
    # H2 Series
    "O1D": "H2D",  # H2D
    "O1E": "H2D Pro",  # H2D Pro
    "O2D": "H2D Pro",  # H2D Pro
    "O1C": "H2C",  # H2C
    "O1C2": "H2C",  # H2C (dual nozzle variant)
    "O1S": "H2S",  # H2S
}

# Serial number prefixes for each model (based on Bambu Lab serial number format)
# Format: MMM??RYMDDUUUUU (15 chars total)
#   MMM = Model prefix (3 chars)
#   ?? = Unknown/revision code (2 chars)
#   R = Revision letter (1 char)
#   Y = Year digit (1 char)
#   M = Month (1 char, hex: 1-9, A=Oct, B=Nov, C=Dec)
#   DD = Day (2 chars)
#   UUUUU = Unit number (5 chars)
MODEL_SERIAL_PREFIXES = {
    # X1 Series
    "BL-P001": "00M00A",  # X1C
    "BL-P002": "00M00A",  # X1
    "C13": "03W00A",  # X1E
    # X2 Series
    "N6": "20P90A",  # X2D (first 4 chars "20P9" match real serials)
    # A2 Series
    "N9": "26A19A",  # A2L (first 5 chars "26A19" match real serials)
    # P Series
    "C11": "01S00A",  # P1P
    "C12": "01P00A",  # P1S
    "N7": "22E00A",  # P2S
    # A1 Series
    "N2S": "03900A",  # A1
    "N1": "03000A",  # A1 Mini
    # H2 Series
    "O1D": "09400A",  # H2D
    "O1E": "09400A",  # H2D Pro (same prefix family as H2D)
    "O2D": "09400A",  # H2D Pro
    "O1C": "09400A",  # H2C
    "O1C2": "09400A",  # H2C (dual nozzle variant)
    "O1S": "09400A",  # H2S
}

# Reverse mapping: display name → SSDP model code (for auto-inheriting from printer model)
DISPLAY_NAME_TO_MODEL_CODE = {v: k for k, v in VIRTUAL_PRINTER_MODELS.items()}

# Default model
DEFAULT_VIRTUAL_PRINTER_MODEL = "BL-P001"  # X1C

# Bound on per-instance ``_slicer_print_options`` cache size. The slicer's
# project_file MQTT command stashes one dict per filename; the
# corresponding ``_add_to_print_queue`` pop only fires when the file
# upload completes. Failed / cancelled / non-3MF uploads orphan their
# stash. The bound triggers FIFO eviction in ``on_print_command`` once
# the dict fills, so a long-running VP can't leak unbounded state.
_SLICER_OPTIONS_CACHE_LIMIT = 128

# How long ``_add_to_print_queue`` waits for the slicer's MQTT
# ``project_file`` after the FTP upload completes (#1780 round 3).
# Bambu Studio sends FTP first, then MQTT immediately after — but on
# wireless / loaded setups the MQTT command can land 2+ s after FTP,
# which used to time the wait out and silently drop ``nozzle_mapping``
# + the other slicer-driven flags. The bumped window covers the
# observed worst case in the field; the late-MQTT fallback in
# ``on_print_command`` covers the rest.
_SLICER_OPTIONS_WAIT_TIMEOUT = 5.0

# How long ``on_print_command`` will retroactively stamp slicer fields
# onto a recently-committed queue item when the MQTT print command
# arrives after ``_SLICER_OPTIONS_WAIT_TIMEOUT`` expired. Covers
# extra-late MQTT (slow wireless slicer, NIC drop+retry) and the
# scheduler tick interval before dispatch picks the item up.
_RECENT_QUEUE_ITEM_TTL = 30.0

# BambuStudio's tri-state calibration options (bed_leveling / flow_cali /
# nozzle_offset_cali) travel on the project_file command as an int — off=0,
# on=1, auto=2 (getValueInt parity). Bed levelling and flow calibration add a
# bool that is true only for "on"; nozzle offset calibration is the int alone.
_TRISTATE_INT = {0: "off", 1: "on", 2: "auto"}


def _tristate_from_slicer(data: dict, bool_field: str | None, int_field: str) -> str | None:
    """Reconstruct off/on/auto from a captured slicer project_file dict.

    Prefer the int companion (auto_bed_leveling / extrude_cali_flag / etc.) which
    carries all three states; fall back to the bool field (on/off only); return
    None when the slicer sent neither so the caller can use its own default.
    """
    if int_field in data:
        try:
            resolved = _TRISTATE_INT.get(int(data[int_field]))
        except (TypeError, ValueError):
            resolved = None
        if resolved is not None:
            return resolved
    if bool_field is not None and bool_field in data:
        return "on" if bool(data[bool_field]) else "off"
    return None


def _extract_slicer_ams_mapping_json(data: dict, log_prefix: str, *, is_dual_nozzle: bool = False) -> str | None:
    """Pull the slicer's own AMS-slot pick out of a captured project_file payload.

    BambuStudio/OrcaSlicer resolves the physical AMS tray for each filament
    live, right before sending — either automatically or via the slicer's
    manual per-filament AMS-slot assignment dialog — and embeds the result as
    ``ams_mapping`` (``list[int]``, position = slot_id-1, value = global tray
    ID) directly in the MQTT ``project_file`` command. Confirmed by wire
    capture: the field is present and already in the exact shape
    ``PrintQueueItem.ams_mapping`` expects.

    The VP-queue path previously never read this — every queued print had the
    scheduler re-derive a mapping from just the 3MF's static type/color at
    dispatch time (`PrintScheduler._compute_ams_mapping_for_printer`), discarding
    the slicer's already-correct, live-resolved pick. That re-derivation can
    land on the wrong physical spool whenever the file's type+color match
    isn't unique (e.g. two spools of the same color) or the file's own
    filament-slot color wasn't what the user actually intended for that
    particular print. Capturing it here — mirroring the existing
    ``nozzle_mapping`` passthrough for H2C rack-swap models (#1780) — lets the
    scheduler's "already resolved, don't touch it" branch in
    ``_ensure_ams_mapping`` use the slicer's own choice unchanged.

    That branch skipping ``_compute_ams_mapping_for_printer`` is also what
    makes this a trade rather than a pure win: ``prefer_lowest_filament``, its
    AMS-filament-backup gate (#1766), the inventory-remain overrides and the
    per-slot force-color overrides all live inside that function. Callers are
    responsible for the gating — this parser only says what the slicer sent.

    An external spool is ``-1`` in the flat list, the same as a filament with
    no tray, and is named only in ``ams_mapping2``. It is resolved from there
    to its global tray (254/255), as the usage capture does (#3166); dispatch
    turns that back into the external-spool entry. Kept as ``-1`` it dispatched
    as unmapped and the printer stopped with 0700-8012 (#3237).
    ``is_dual_nozzle`` is the target printer's, since that decides which
    global tray an external spool is.

    Returns ``None`` when the field is absent, unparsable, or the classic
    "all -1" unresolved-race sentinel (#2589) — never worth trusting over a
    fresh live computation.
    """
    raw = data.get("ams_mapping")
    if raw is None:
        return None
    if isinstance(raw, str):
        try:
            raw = json.loads(raw)
        except json.JSONDecodeError:
            logger.warning("%s Slicer ams_mapping is unparseable JSON, dropping: %r", log_prefix, raw)
            return None
    # bool is a subclass of int in Python — isinstance(True, int) is True —
    # so it must be excluded explicitly, or [True, False] would pass as a
    # valid mapping.
    if not isinstance(raw, list) or not raw or not all(isinstance(v, int) and not isinstance(v, bool) for v in raw):
        return None
    detail = data.get("ams_mapping2")
    if isinstance(detail, str):
        try:
            detail = json.loads(detail)
        except json.JSONDecodeError:
            detail = None
    raw = resolve_external_spools_in_mapping(raw, detail, is_dual_nozzle)
    if all(v < 0 for v in raw):
        # #2589 sentinel — every slot unresolved. Let the scheduler compute a
        # fresh mapping from live AMS state instead of trusting this.
        return None
    return json.dumps(raw)


def _get_serial_for_model(model: str, serial_suffix: str) -> str:
    """Get serial number for the given model and suffix."""
    prefix = MODEL_SERIAL_PREFIXES.get(model, "00M09A")
    return f"{prefix}{serial_suffix}"


class VirtualPrinterInstance:
    """Per-printer state and file handling logic.

    Each instance represents one virtual printer with its own config,
    upload directory, certificates, and file handling mode.
    """

    def __init__(
        self,
        *,
        vp_id: int,
        name: str,
        mode: str,
        model: str,
        access_code: str,
        serial_suffix: str,
        target_printer_ip: str = "",
        target_printer_serial: str = "",
        target_printer_id: int | None = None,
        auto_dispatch: bool = True,
        queue_force_color_match: bool = False,
        save_ams_mapping: bool = False,
        gcode_injection: bool = False,
        bind_ip: str = "",
        remote_interface_ip: str = "",
        tailscale_disabled: bool = True,
        base_dir: Path,
        session_factory: Callable | None = None,
        printer_manager: "PrinterManager | None" = None,
    ):
        self.id = vp_id
        self.name = name
        # Normalize on construction so the rest of the code only compares
        # canonical values, even when a legacy DB row hasn't been migrated
        # yet (e.g. fresh-from-disk during the boot window before the
        # one-shot migration in `core/database.py` has executed).
        self.mode = normalize_vp_mode(mode) or VP_MODE_ARCHIVE
        self.model = model
        self.access_code = access_code
        self.serial_suffix = serial_suffix
        self.target_printer_ip = target_printer_ip
        self.target_printer_serial = target_printer_serial
        self.target_printer_id = target_printer_id
        self.auto_dispatch = auto_dispatch
        self.queue_force_color_match = queue_force_color_match
        self.save_ams_mapping = save_ams_mapping
        self.gcode_injection = gcode_injection
        self.bind_ip = bind_ip
        self.remote_interface_ip = remote_interface_ip
        self.tailscale_disabled = tailscale_disabled
        self._session_factory = session_factory
        self._printer_manager = printer_manager

        # Directories
        self.upload_dir = base_dir / "uploads" / str(vp_id)
        self.cert_dir = base_dir / "certs" / str(vp_id)
        shared_ca_dir = base_dir / "certs"

        # Ensure directories exist
        self.upload_dir.mkdir(parents=True, exist_ok=True)
        (self.upload_dir / "cache").mkdir(exist_ok=True)
        self.cert_dir.mkdir(parents=True, exist_ok=True)

        # Certificate service (shared CA, per-instance printer cert)
        self._cert_service = CertificateService(
            cert_dir=self.cert_dir,
            serial=self.serial,
            shared_ca_dir=shared_ca_dir,
        )

        # Pending files for MQTT correlation
        self._pending_files: dict[str, Path] = {}

        # Slicer-side print options captured from the MQTT `project_file`
        # command, keyed by filename. Used by `_add_to_print_queue` so the
        # queue item inherits the user's slicer-chosen timelapse / bed_leveling
        # / flow_cali / vibration_cali / layer_inspect / use_ams toggles rather
        # than falling back to the global `default_*` settings (#1403). FTP
        # completes a few hundred ms before the slicer's MQTT `project_file`
        # arrives, so the queue-add path waits briefly on the event below
        # before reading the dict. Events are popped along with the options
        # so the dict stays bounded.
        self._slicer_print_options: dict[str, dict] = {}
        self._slicer_print_options_events: dict[str, asyncio.Event] = {}

        # Queue items recently committed by `_add_to_print_queue`, keyed by
        # FTP filename. Used by `on_print_command` to retroactively stamp the
        # slicer's nozzle_mapping (and the other slicer-driven flags) onto a
        # queue item when the MQTT `project_file` arrives after the queue-add
        # wait timed out — the #1780 round-3 race. Value is
        # (queue_item_ids, monotonic_committed_at); entries older than
        # `_RECENT_QUEUE_ITEM_TTL` are evicted opportunistically on each
        # queue-add.
        self._recent_queue_items: dict[str, tuple[list[int], float]] = {}
        # Per entry above: the highest slot id each queued plate prints, so a
        # late slicer AMS mapping is stamped only onto plates it reaches.
        self._recent_queue_item_reach: dict[str, dict[int, int]] = {}

        # Per-instance services
        self._proxy: SlicerProxyManager | None = None
        self._ftp: VirtualPrinterFTPServer | None = None
        self._mqtt: SimpleMQTTServer | None = None
        self._mqtt_bridge: MQTTBridge | None = None
        self._rtsp_proxy: TCPProxy | None = None
        self._bind: BindServer | None = None
        self._ssdp: VirtualPrinterSSDPServer | None = None
        self._ssdp_proxy: SSDPProxy | None = None
        self._tasks: list[asyncio.Task] = []

        # Pending timer that re-fires gcode_state=FINISH after a project_file
        # ack. See ``_schedule_finish_release`` for the #1658 rationale.
        self._finish_release_task: asyncio.Task | None = None

    @property
    def serial(self) -> str:
        """Full serial number for this virtual printer."""
        return _get_serial_for_model(self.model or DEFAULT_VIRTUAL_PRINTER_MODEL, self.serial_suffix)

    @property
    def cert_path(self) -> Path:
        return self._cert_service.cert_path

    @property
    def key_path(self) -> Path:
        return self._cert_service.key_path

    @property
    def is_proxy(self) -> bool:
        return self.mode == "proxy"

    @property
    def is_running(self) -> bool:
        return len(self._tasks) > 0 and all(not t.done() for t in self._tasks)

    def generate_certificates(self) -> tuple[Path, Path]:
        """Generate certificates for this instance."""
        self._cert_service.serial = self.serial if not self.is_proxy else (self.target_printer_serial or self.serial)
        additional_ips = [self.remote_interface_ip] if self.remote_interface_ip else None
        if self.bind_ip:
            additional_ips = additional_ips or []
            additional_ips.append(self.bind_ip)
        self._cert_service.delete_printer_certificate()
        return self._cert_service.generate_certificates(additional_ips=additional_ips)

    # -- File handling callbacks --

    async def on_file_received(self, file_path: Path, source_ip: str) -> None:
        """Handle file upload completion from FTP."""
        logger.info("[VP %s] Received file: %s from %s", self.name, file_path.name, source_ip)

        self._pending_files[file_path.name] = file_path

        # Accept both canonical (`archive`/`queue`) and legacy
        # (`immediate`/`print_queue`) wire values so a stale row that hasn't
        # been migrated yet still dispatches correctly. Migration in
        # `core/database.py` rewrites existing rows once at boot.
        mode = normalize_vp_mode(self.mode)
        if mode == VP_MODE_ARCHIVE:
            await self._archive_file(file_path, source_ip)
        elif mode == VP_MODE_QUEUE:
            await self._add_to_print_queue(file_path, source_ip)
        else:
            await self._queue_file(file_path, source_ip)

        # Signal job completion to the slicer. Send-flow slicers don't watch the
        # post-upload state and would be happy with anything; the Print flow
        # (intended for proxy-mode VPs, but users sometimes click it against
        # queue/immediate/review modes too — #1280) watches the gcode_state
        # cycle and only releases its in-flight-job lock when it sees FINISH.
        # Going PREPARE → IDLE wedges the slicer's UI at "Downloading...(0%)"
        # and blocks the next dispatch with "busy with another print job".
        # PREPARE → FINISH satisfies both flows. prepare_percent=100 also
        # unfreezes the slicer's "Downloading X%" progress bar which it ticks
        # against the same field during the upload window.
        if self._mqtt and file_path.suffix.lower() == ".3mf":
            self._mqtt.set_gcode_state("FINISH", filename=file_path.name, prepare_percent="100")
            # FINISH is the terminal state for the upload cycle per #1280
            # (commit 0d6171dc). The Print-flow slicer's in-flight-job lock
            # releases on FINISH; resetting to IDLE 2 s later would re-confuse
            # the slicer that just unwedged. Earlier audit suggesting the
            # IDLE reset was wrong — staying at FINISH is the designed
            # behaviour. The next upload's PREPARE→FINISH cycle starts fresh.

    def _keeps_slicer_nozzle_mapping(self) -> bool:
        """Whether the slicer's ``nozzle_mapping`` is stored on the queue item.

        On a rack model (H2C) it holds physical rack positions, which name the
        rack of the printer the slicer was looking at: kept only when this VP
        targets that one printer. A VP for any printer of the model leaves it
        empty, and the dispatch resolves the positions against the rack of the
        printer it picks. Elsewhere it holds extruder indices, which are the
        same on every printer of the model.
        """
        return self.target_printer_id is not None or not is_nozzle_rack_model(self.model)

    def _keeps_slicer_use_ams(self) -> bool:
        """Whether the slicer's ``use_ams`` is stored on the queue item.

        The slicer was looking at this VP, which advertises no AMS. Only a VP
        bound to one printer stands for a machine whose AMS the user chose to
        use or not; a VP for any printer of the model leaves it on and the
        dispatch reconciles it against the printer it picks (#2595).
        """
        return self.target_printer_id is not None

    def _target_is_dual_nozzle(self) -> bool:
        """Whether the target printer has two nozzles, which decides the global
        tray of an external spool in the slicer's mapping (#3237).

        Same signal dispatch uses: the live client's detection, else its model.
        Without a client, the VP's own model is what the slicer sliced for.
        """
        from backend.app.utils.printer_models import is_dual_nozzle_model

        client = (
            self._printer_manager.get_client(self.target_printer_id)
            if self._printer_manager is not None and self.target_printer_id is not None
            else None
        )
        if client is not None:
            return bool(getattr(client, "_is_dual_nozzle", False)) or is_dual_nozzle_model(
                getattr(client, "model", None)
            )
        return is_dual_nozzle_model(self.model)

    async def on_print_command(self, filename: str, data: dict) -> None:
        """Handle print command from MQTT.

        Captures the slicer's project_file options (`timelapse`, `bed_leveling`,
        `flow_cali`, `vibration_cali`, `layer_inspect`, `use_ams`, plus the
        H2C rack-pick `nozzle_mapping`) so the VP-queue path can inherit them
        when adding the item to the queue, rather than falling back to the
        global default settings (#1403, #1780).
        Only queue mode consumes the capture; archive / review / proxy
        modes ignore the print command, so we skip the stash there to keep
        the dict from accumulating one entry per print over the VP's
        uptime.

        Also schedules the #1658 follow-up that re-fires gcode_state=FINISH a
        moment after the synthetic project_file ack — for every non-proxy
        mode — so the slicer's "Downloading" UI releases on the slicer's
        FTP-first-then-MQTT send order.

        ``filename`` is the slicer's ``subtask_name`` (bare model name, no
        extension) — used verbatim for `_schedule_finish_release` because
        push_status echoes it back to the slicer as gcode_file / subtask_name.
        The queue-side stash key is derived from ``data["file"]`` (the FTP
        filename with extension) so `_add_to_print_queue`'s
        ``file_path.name`` lookup matches; falls back to ``filename`` when
        ``data["file"]`` is absent (legacy slicers / non-3MF uploads).
        Stash/lookup mismatch was the #1780 root cause — every captured field
        silently fell back to settings defaults on every Bambu Studio "Send".
        """
        logger.info("[VP %s] Print command for: %s", self.name, filename)
        mode = normalize_vp_mode(self.mode)
        if mode != VP_MODE_PROXY and filename and self._mqtt is not None:
            self._schedule_finish_release(filename)
        if mode != VP_MODE_QUEUE:
            return
        # Stash key must match `_add_to_print_queue`'s lookup, which uses
        # `file_path.name` (FTP filename WITH extension). The slicer's
        # `subtask_name` (== this method's `filename` arg) is the bare model
        # name, no extension — using it as the stash key was the #1780 root
        # cause.
        stash_key = data.get("file") or filename
        # Drop the oldest stash if the cache is growing — happens when the
        # slicer sends project_file for a filename whose FTP upload was
        # rejected / cancelled / non-3MF, so _add_to_print_queue's pop
        # never fires. With no bound, a long-running VP accumulates one
        # dict per such mismatch.
        if len(self._slicer_print_options) >= _SLICER_OPTIONS_CACHE_LIMIT:
            try:
                stale_key = next(iter(self._slicer_print_options))
                self._slicer_print_options.pop(stale_key, None)
                self._slicer_print_options_events.pop(stale_key, None)
                logger.debug("[VP %s] Evicted stale slicer options for %s", self.name, stale_key)
            except StopIteration:
                pass
        self._slicer_print_options[stash_key] = dict(data)
        event = self._slicer_print_options_events.get(stash_key)
        if event:
            event.set()
            return
        # No consumer waiting: `_add_to_print_queue` either already gave up
        # (wait_for timed out) or hasn't started yet (FTP still uploading).
        # If a queue item was committed within the last
        # `_RECENT_QUEUE_ITEM_TTL`, the wait timed out and the row holds
        # settings defaults instead of the slicer's pick — retroactively
        # stamp the slicer-driven fields so the dispatcher honours the
        # user's choice. Covers the #1780 round-3 race where Bambu Studio's
        # MQTT lands just past the bumped wait ceiling.
        await self._restamp_recent_queue_item(stash_key, data)

    async def _restamp_recent_queue_item(self, stash_key: str, data: dict) -> None:
        """Patch slicer-driven fields onto a queue item the MQTT command missed.

        ``_add_to_print_queue`` waits up to ``_SLICER_OPTIONS_WAIT_TIMEOUT``
        for the slicer's MQTT ``project_file`` before committing the queue
        item. If the MQTT command arrives after that window — observed in
        the field at ~2.1 s on H2C / wireless setups (#1780 round 3) — the
        row was already written with settings defaults. This method runs
        on the late MQTT path: it looks up the most recent queue items
        committed for this filename and patches in the slicer's
        ``nozzle_mapping`` + ``ams_mapping`` + workflow flags, but only
        while the items are still ``pending`` (scheduler hasn't dispatched
        them yet). The two mappings pass the same gates as in
        ``_add_to_print_queue``.
        """
        if not self._session_factory:
            return
        entry = self._recent_queue_items.get(stash_key)
        plate_reach = self._recent_queue_item_reach.pop(stash_key, {})
        if entry is None:
            return
        queue_item_ids, committed_at = entry
        if time.monotonic() - committed_at > _RECENT_QUEUE_ITEM_TTL:
            self._recent_queue_items.pop(stash_key, None)
            return

        import json

        # Mirror the field set `_add_to_print_queue` reads off slicer_opts.
        # MQTT uses `bed_leveling` (single L); the column is `bed_levelling`.
        # `nozzles_info` is intentionally not stamped — column kept for
        # legacy rows but never written; see PrintQueueItem.nozzles_info.
        patch: dict = {}
        # Tri-state options (off/on/auto) — reconstruct from the int companion.
        for bool_field, int_field, column in (
            ("bed_leveling", "auto_bed_leveling", "bed_levelling"),
            ("flow_cali", "extrude_cali_flag", "flow_cali"),
            (None, "nozzle_offset_cali", "nozzle_offset_cali"),
        ):
            resolved = _tristate_from_slicer(data, bool_field, int_field)
            if resolved is not None:
                patch[column] = resolved
        # On/off options.
        for mqtt_field, column in (
            ("vibration_cali", "vibration_cali"),
            ("layer_inspect", "layer_inspect"),
            ("timelapse", "timelapse"),
            ("use_ams", "use_ams"),
        ):
            if mqtt_field == "use_ams" and not self._keeps_slicer_use_ams():
                continue
            if mqtt_field in data:
                patch[column] = bool(data[mqtt_field])

        # Same gate as in `_add_to_print_queue`.
        raw = data.get("nozzle_mapping") if self._keeps_slicer_nozzle_mapping() else None
        if raw is not None:
            if isinstance(raw, str):
                try:
                    raw = json.loads(raw)
                except json.JSONDecodeError:
                    logger.warning(
                        "[VP %s] Late MQTT nozzle_mapping is unparseable JSON, dropping: %r",
                        self.name,
                        raw,
                    )
                    raw = None
            if raw is not None:
                patch["nozzle_mapping"] = json.dumps(raw)

        # Same two gates as the immediate path in `_add_to_print_queue`: a
        # model-based VP has no live AMS layout for the slicer to have resolved
        # tray IDs against, and taking the slicer's pick at all is the per-VP
        # `save_ams_mapping` opt-in (it makes the scheduler skip
        # `_compute_ams_mapping_for_printer`, and with it prefer-lowest and the
        # #1766 backup gate).
        ams_mapping_json = (
            _extract_slicer_ams_mapping_json(
                data, f"[VP {self.name}] Late MQTT", is_dual_nozzle=self._target_is_dual_nozzle()
            )
            if self.target_printer_id is not None and self.save_ams_mapping
            else None
        )
        # `Force color match` still wins for this dispatch — see the same
        # decision in `_add_to_print_queue`. The archive patch below is
        # deliberately not gated on it: persisting the pick for later reprints
        # is exactly what the toggle promises.
        if ams_mapping_json is not None and not self.queue_force_color_match:
            patch["ams_mapping"] = ams_mapping_json

        # `ams_mapping_json` alone is enough to keep going even when `patch` is
        # empty: with `Force color match` on it never reaches the queue item,
        # but it still has to be written onto the archive below.
        if not patch and ams_mapping_json is None:
            self._recent_queue_items.pop(stash_key, None)
            return

        from sqlalchemy import select, update

        from backend.app.models.archive import PrintArchive
        from backend.app.models.print_queue import PrintQueueItem

        try:
            async with self._session_factory() as db:
                # Only stamp items still pending; once the scheduler has
                # picked the row up we can't safely race the dispatcher.
                result = await db.execute(
                    select(PrintQueueItem.id, PrintQueueItem.archive_id).where(
                        PrintQueueItem.id.in_(queue_item_ids),
                        PrintQueueItem.status == "pending",
                    )
                )
                rows = result.all()
                eligible_ids = [row[0] for row in rows]
                if not eligible_ids:
                    self._recent_queue_items.pop(stash_key, None)
                    return
                # As in `_add_to_print_queue`: a mapping shorter than a plate's
                # highest slot id cannot address that plate's own slots, so that
                # plate is left to the scheduler's computed mapping.
                short_ids: list[int] = []
                if "ams_mapping" in patch:
                    reach = len(json.loads(patch["ams_mapping"]))
                    short_ids = [i for i in eligible_ids if plate_reach.get(i, 0) > reach]
                    if short_ids:
                        logger.warning(
                            "[VP %s] Late slicer ams_mapping has %d entries, too few for queue item(s) %s; "
                            "leaving those to the scheduler's mapping.",
                            self.name,
                            reach,
                            short_ids,
                        )
                reached_ids = [i for i in eligible_ids if i not in short_ids]
                if patch and reached_ids:
                    await db.execute(update(PrintQueueItem).where(PrintQueueItem.id.in_(reached_ids)).values(**patch))
                short_patch = {k: v for k, v in patch.items() if k != "ams_mapping"}
                if short_ids and short_patch:
                    await db.execute(
                        update(PrintQueueItem).where(PrintQueueItem.id.in_(short_ids)).values(**short_patch)
                    )

                # The archive was already created (with no slicer_ams_mapping)
                # before this late MQTT arrived — see
                # `_extract_slicer_ams_mapping_json`'s docstring. Patch it here
                # too so a reprint later still picks up the slicer's pick, and
                # the "AMS mapping from slicer" badge reflects reality instead
                # of staying stuck on the archive's initial (empty) snapshot.
                # Already gated on `save_ams_mapping` above, and deliberately
                # NOT on `queue_force_color_match`: that toggle decides how
                # *this* print is matched, not whether the pick is worth
                # keeping for a later reprint.
                if ams_mapping_json is not None:
                    archive_ids = {row[1] for row in rows if row[1] is not None}
                    if archive_ids:
                        archive_result = await db.execute(select(PrintArchive).where(PrintArchive.id.in_(archive_ids)))
                        for archive in archive_result.scalars().all():
                            extra = dict(archive.extra_data or {})
                            extra["slicer_ams_mapping"] = {
                                "mapping": json.loads(ams_mapping_json),
                                "printer_id": self.target_printer_id,
                            }
                            archive.extra_data = extra

                await db.commit()
                logger.info(
                    "[VP %s] Late slicer MQTT for %s — retroactively stamped %s onto queue item(s) %s%s",
                    self.name,
                    stash_key,
                    sorted(patch.keys()),
                    eligible_ids,
                    " and saved the slicer's AMS pick onto the archive" if ams_mapping_json is not None else "",
                )
        except Exception as e:
            logger.error(
                "[VP %s] Failed to retroactively stamp queue item(s) %s for %s: %s",
                self.name,
                queue_item_ids,
                stash_key,
                e,
            )
        finally:
            self._recent_queue_items.pop(stash_key, None)

    def _schedule_finish_release(self, filename: str, delay: float = 1.5) -> None:
        """Re-set gcode_state=FINISH on the VP after the project_file ack.

        #1280 set FINISH after the FTP upload completes — that was correct
        for the slicer flow at the time (MQTT project_file → FTP → done).
        Bambu Studio 2.7.x flipped the order to FTP → FTP → MQTT project_file,
        which means ``_send_print_response`` runs *after* the FINISH set in
        ``on_file_received`` and overwrites the state back to PREPARE. The
        slicer's 1 Hz status stream then carries PREPARE forever and the
        send modal sits at "Downloading" until the VP is restarted (#1658).

        Re-firing FINISH after a short delay closes the gap: the slicer sees
        the synthetic PREPARE in the project_file ack (and likely one PREPARE
        push on the 1 Hz cycle), then the next push carries FINISH and the
        modal releases. Proxy mode is exempt — there the real printer drives
        the state through the bridge and a synthetic FINISH would clobber a
        real PREPARE/RUNNING transition coming back from the printer.

        Cancels any in-flight timer before scheduling a new one so a slicer
        that fires project_file twice in quick succession only ends in one
        FINISH.
        """
        if self._mqtt is None:
            return
        if self._finish_release_task is not None and not self._finish_release_task.done():
            self._finish_release_task.cancel()
        self._finish_release_task = asyncio.create_task(
            self._delayed_finish_release(filename, delay),
            name=f"vp-{self.id}-finish-release",
        )

    async def _delayed_finish_release(self, filename: str, delay: float) -> None:
        """Sleep, then set gcode_state=FINISH. Used by ``_schedule_finish_release``."""
        try:
            await asyncio.sleep(delay)
        except asyncio.CancelledError:
            return
        if self._mqtt is None:
            return
        self._mqtt.set_gcode_state("FINISH", filename=filename, prepare_percent="100")
        logger.debug("[VP %s] Re-set gcode_state=FINISH after project_file ack (%s)", self.name, filename)

    async def _archive_file(self, file_path: Path, source_ip: str) -> None:
        """Archive file immediately."""
        if not self._session_factory:
            logger.error("Cannot archive: no database session factory configured")
            return

        if file_path.suffix.lower() != ".3mf":
            logger.debug("Skipping non-3MF file: %s", file_path.name)
            self._pending_files.pop(file_path.name, None)
            try:
                file_path.unlink()
            except OSError:
                pass
            return

        archived = False
        try:
            from backend.app.api.routes.settings import get_setting
            from backend.app.services.archive import ArchiveService

            async with self._session_factory() as db:
                name_source = await get_setting(db, "virtual_printer_archive_name_source")
                prefer_filename = name_source == "filename"
                service = ArchiveService(db)
                archive = await service.archive_print(
                    printer_id=None,
                    source_file=file_path,
                    print_data={
                        "status": "archived",
                        "source": "virtual_printer",
                        "source_ip": source_ip,
                    },
                    prefer_filename_for_name=prefer_filename,
                )
                if archive:
                    logger.info("[VP %s] Archived: %s - %s", self.name, archive.id, archive.print_name)
                    await self._broadcast_archive_created(archive)
                    archived = True
                else:
                    logger.error("Failed to archive file: %s", file_path.name)
        except Exception as e:
            logger.error("Error archiving file: %s", e)
        finally:
            # Always release the in-flight marker and delete the temp file —
            # previously the failure paths only logged and the next upload of
            # the same name was silently rejected with "already uploading",
            # the upload_dir filled up indefinitely, and the slicer received
            # a clean 226 even though no archive existed (#audit-R2-1).
            self._pending_files.pop(file_path.name, None)
            if archived:
                try:
                    file_path.unlink()
                except OSError:
                    pass
            else:
                # Drop the failed temp file so it doesn't accumulate.
                try:
                    file_path.unlink(missing_ok=True)
                except OSError:
                    pass

    async def _queue_file(self, file_path: Path, source_ip: str) -> None:
        """Queue file for user review."""
        if not self._session_factory:
            logger.error("Cannot queue: no database session factory configured")
            return

        if file_path.suffix.lower() != ".3mf":
            self._pending_files.pop(file_path.name, None)
            try:
                file_path.unlink()
            except OSError:
                pass
            return

        # Peek at the 3MF for the embedded title BEFORE we hand it off to the
        # DB. Storing it now means the /pending-uploads/ list doesn't have to
        # reopen every 3MF on every render to keep the review card and the
        # eventual archive name in sync (#1152 follow-up). Failure to parse is
        # not fatal — the response model falls back to the filename stem.
        metadata_print_name: str | None = None
        try:
            from backend.app.services.archive import ThreeMFParser

            parsed = ThreeMFParser(file_path).parse()
            raw_name = parsed.get("print_name")
            if isinstance(raw_name, str) and raw_name.strip():
                metadata_print_name = raw_name.strip()[:255]
        except Exception as e:
            logger.debug("[VP %s] Metadata title peek failed for %s: %s", self.name, file_path.name, e)

        try:
            from backend.app.models.pending_upload import PendingUpload

            async with self._session_factory() as db:
                pending = PendingUpload(
                    filename=file_path.name,
                    file_path=str(file_path),
                    file_size=file_path.stat().st_size,
                    source_ip=source_ip,
                    status="pending",
                    uploaded_at=datetime.now(timezone.utc),
                    metadata_print_name=metadata_print_name,
                )
                db.add(pending)
                await db.commit()
                logger.info("[VP %s] Queued: %s - %s", self.name, pending.id, file_path.name)
        except Exception as e:
            logger.error("Error queueing file: %s", e)
            # Queue insert failed — drop the temp file so it doesn't
            # accumulate. The file is unreachable without the DB row.
            try:
                file_path.unlink(missing_ok=True)
            except OSError:
                pass
        finally:
            # Always release the in-flight marker so concurrent uploads
            # with the same filename aren't spuriously rejected after
            # a queue failure.
            self._pending_files.pop(file_path.name, None)

    async def _add_to_print_queue(self, file_path: Path, source_ip: str) -> None:
        """Archive file and add to print queue, assigned to target printer or model."""
        if not self._session_factory:
            logger.error("Cannot add to print queue: no database session factory configured")
            return

        if file_path.suffix.lower() != ".3mf":
            self._pending_files.pop(file_path.name, None)
            try:
                file_path.unlink()
            except OSError:
                pass
            return

        # Wait briefly for the slicer's MQTT `project_file` command so the
        # queue item can inherit the slicer-side print options the user
        # picked (timelapse, bed_leveling, etc). Slicers send the FTP upload
        # first and the MQTT command immediately after, so the typical lag
        # is a few hundred ms. The window is generous enough to absorb
        # wireless / loaded-Pi jitter without making every VP-queue add
        # visibly slow — observed worst case in #1780 round 3 was 2.085 s,
        # the previous 2.0 s ceiling. Falls back to the global default_*
        # settings if MQTT doesn't arrive in time (legacy behaviour for
        # users on a slicer that doesn't send a print command). #1403.
        # The wait is skipped when there's no MQTT server attached — covers
        # unit tests that invoke `_add_to_print_queue` directly without
        # going through `on_print_command`, so they don't pay the wait tax.
        slicer_opts = self._slicer_print_options.pop(file_path.name, None)
        if slicer_opts is None and self._mqtt is not None:
            event = asyncio.Event()
            self._slicer_print_options_events[file_path.name] = event
            try:
                await asyncio.wait_for(event.wait(), timeout=_SLICER_OPTIONS_WAIT_TIMEOUT)
                slicer_opts = self._slicer_print_options.pop(file_path.name, None)
            except TimeoutError:
                slicer_opts = None
            finally:
                self._slicer_print_options_events.pop(file_path.name, None)
        # If the cache still misses, queued workflow flags / nozzle pick will
        # silently fall back to settings defaults. Surface the missed key so a
        # future stash/lookup mismatch (the #1780 root cause) is obvious in
        # the log instead of needing a wire capture to diagnose.
        if slicer_opts is None:
            logger.debug(
                "[VP %s] No slicer options cached for %r (cache keys: %s); "
                "workflow flags + nozzle pick will fall back to settings defaults.",
                self.name,
                file_path.name,
                sorted(self._slicer_print_options.keys()),
            )

        try:
            import json

            from backend.app.api.routes.settings import get_setting
            from backend.app.models.print_queue import PrintQueueItem
            from backend.app.services.archive import ArchiveService
            from backend.app.services.filament_requirements import extract_filament_requirements
            from backend.app.services.print_confirmation import confirm_outcome_for_new_queue_item
            from backend.app.services.print_defaults import default_print_options

            async with self._session_factory() as db:
                name_source = await get_setting(db, "virtual_printer_archive_name_source")
                prefer_filename = name_source == "filename"

                # Workflow defaults from settings. Without this the
                # PrintQueueItem below would fall back to the column-level
                # defaults and ignore the user's workflow preferences (#1235).
                # The slicer-side options captured above (if any) take
                # precedence per-field over these defaults.
                defaults = await default_print_options(db)

                def _slicer_or(field_mqtt: str, settings_default: bool) -> bool:
                    """Slicer's MQTT value if present, else the settings default.

                    Slicer payloads carry both bool and int (0/1) shapes
                    depending on firmware family — coerce via bool() so
                    `0`/`False` and `1`/`True` both work.
                    """
                    if slicer_opts is not None and field_mqtt in slicer_opts:
                        return bool(slicer_opts[field_mqtt])
                    return settings_default

                def _slicer_tristate(bool_field: str | None, int_field: str, settings_default: str) -> str:
                    """Slicer's tri-state (off/on/auto) if present, else the default."""
                    if slicer_opts is not None:
                        resolved = _tristate_from_slicer(slicer_opts, bool_field, int_field)
                        if resolved is not None:
                            return resolved
                    return settings_default

                # Note the MQTT field names differ from Bambuddy's column
                # names: MQTT uses `bed_leveling` (single L) while the
                # column / settings key use `bed_levelling` (double L).
                bed_levelling = _slicer_tristate(
                    "bed_leveling",
                    "auto_bed_leveling",
                    defaults["bed_levelling"],
                )
                flow_cali = _slicer_tristate(
                    "flow_cali",
                    "extrude_cali_flag",
                    defaults["flow_cali"],
                )
                nozzle_offset_cali = _slicer_tristate(
                    None,
                    "nozzle_offset_cali",
                    defaults["nozzle_offset_cali"],
                )
                vibration_cali = _slicer_or("vibration_cali", defaults["vibration_cali"])
                layer_inspect = _slicer_or("layer_inspect", defaults["layer_inspect"])
                timelapse = _slicer_or("timelapse", defaults["timelapse"])
                # No setting for it: the column default (on) when the slicer is
                # silent, or when the VP does not stand for one printer.
                use_ams = _slicer_or("use_ams", True) if self._keeps_slicer_use_ams() else True

                # "Ask for Outcome" is a default print option like the ones
                # above. A plate sent here from Bambu Studio is also one of the
                # prints `confirm_outcome_external_prints` names — but it
                # arrives with a queue item, so on_print_start resumes the
                # archive created below instead of treating it as external, and
                # without this the setting could never reach it (#1898).
                confirm_outcome = await confirm_outcome_for_new_queue_item(db, started_outside_bambuddy=True)

                # H2C dual-nozzle-rack slicer-pick preservation (#1780).
                # BambuStudio's project_file MQTT command for rack-swap models
                # (O1C2 today) carries `nozzle_mapping` — a per-filament array
                # of physical nozzle position IDs (`list[int]`). Forward it
                # verbatim onto the queue item so the dispatcher can replay it
                # in its own project_file command. Without this the H2C
                # firmware falls back to "last matching nozzle" auto-pick and
                # ignores the user's Bambu Studio choice. Every other model
                # has it absent from slicer_opts, so the capture is a
                # transparent no-op there. (`nozzles_info` was also captured
                # in the original fix but BambuStudio never actually sends it
                # — verified via wire capture on H2C — so only `nozzle_mapping`
                # is forwarded now.)
                #
                # On a rack model only for a VP that targets one printer: the
                # positions name that printer's rack. An "Any <model>" VP leaves
                # them to the dispatch, which resolves them against the rack of
                # the printer it picks. See `_keeps_slicer_nozzle_mapping`.
                nozzle_mapping_json: str | None = None
                if slicer_opts is not None and self._keeps_slicer_nozzle_mapping():
                    raw = slicer_opts.get("nozzle_mapping")
                    if raw is not None:
                        # BambuStudio's NetworkAgent embeds this as parsed
                        # JSON in the project_file body (matching the
                        # ams_mapping shape Bambuddy already consumes as
                        # list[int]). Accept a JSON-encoded string defensively
                        # in case any path arrives stringified.
                        if isinstance(raw, str):
                            try:
                                raw = json.loads(raw)
                            except json.JSONDecodeError:
                                logger.warning(
                                    "[VP %s] Slicer nozzle_mapping is unparseable JSON, dropping: %r",
                                    self.name,
                                    raw,
                                )
                                raw = None
                        if raw is not None:
                            nozzle_mapping_json = json.dumps(raw)

                # Slicer's own live-resolved AMS-slot pick (see docstring on
                # `_extract_slicer_ams_mapping_json`). Stamped onto every plate
                # below, same treatment as nozzle_mapping_json above — when
                # present it makes `_ensure_ams_mapping` skip its own
                # type/color re-derivation entirely and dispatch use exactly
                # the tray the slicer/user picked.
                #
                # Two gates, both required:
                #
                # 1. This VP must target one fixed printer. A model-based
                #    ("Any <model>") VP has no MQTT bridge to a real printer,
                #    so the slicer has no live AMS layout to resolve tray IDs
                #    against — whatever it sends here is meaningless (or,
                #    worse, coincidentally valid for the wrong printer once
                #    the scheduler later picks one).
                # 2. The per-VP `save_ams_mapping` opt-in must be on. Taking
                #    the slicer's pick means `_ensure_ams_mapping` returns
                #    early and `_compute_ams_mapping_for_printer` never runs —
                #    and that function is where `prefer_lowest_filament`, its
                #    AMS-filament-backup gate (#1766) and the inventory-remain
                #    overrides live. Honouring the slicer unconditionally would
                #    silently retire all of that for every existing queue-mode
                #    VP on upgrade, so it's opt-in like every other queue-mode
                #    behaviour toggle (#2700 review).
                #
                # Either gate failing leaves it unset, and the scheduler's
                # normal type/color re-derivation runs against whichever
                # printer actually gets the job.
                ams_mapping_json: str | None = None
                if slicer_opts is not None and self.target_printer_id is not None and self.save_ams_mapping:
                    ams_mapping_json = _extract_slicer_ams_mapping_json(
                        slicer_opts, f"[VP {self.name}]", is_dual_nozzle=self._target_is_dual_nozzle()
                    )

                # `Force color match` is the user asking Bambuddy to do the
                # matching strictly, against the printer's live trays. Its only
                # effect on a fixed-printer item is via the per-slot
                # `filament_overrides` written below, which are consumed inside
                # `_compute_ams_mapping_for_printer` — the exact function a
                # stored mapping skips. So when both toggles are on, the
                # explicit strictness wins for *this* dispatch and the slicer's
                # pick is still persisted onto the archive for later reprints,
                # which is what `Save AMS mapping` actually promises (#2700
                # review).
                queue_ams_mapping_json = ams_mapping_json
                if queue_ams_mapping_json is not None and self.queue_force_color_match:
                    logger.info(
                        "[VP %s] Saved the slicer's AMS pick to the archive but not onto the queue item(s): "
                        "'Force color match' is on, so the scheduler matches against live trays for this print.",
                        self.name,
                    )
                    queue_ams_mapping_json = None

                # Parsed once for the per-plate length check in the loop below.
                queue_ams_mapping = json.loads(queue_ams_mapping_json) if queue_ams_mapping_json else None

                service = ArchiveService(db)
                archive = await service.archive_print(
                    printer_id=None,
                    source_file=file_path,
                    print_data={
                        "status": "archived",
                        "source": "virtual_printer",
                        "source_ip": source_ip,
                    },
                    prefer_filename_for_name=prefer_filename,
                    # Slicer's own live AMS-slot pick -- promoted to
                    # `extra_data.slicer_ams_mapping` by archive_print() so a
                    # later reprint can reuse it. Already gated on the per-VP
                    # `save_ams_mapping` opt-in above. Tagged with the printer
                    # it was resolved against so a later reprint on a
                    # *different* printer knows not to reuse it (#2700 review).
                    slicer_ams_mapping=(json.loads(ams_mapping_json) if ams_mapping_json else None),
                    slicer_ams_mapping_printer_id=self.target_printer_id,
                )
                if archive:
                    logger.info("[VP %s] Archived: %s - %s", self.name, archive.id, archive.print_name)
                    # Assign to specific printer if configured, otherwise use model for "Any X" scheduling
                    target_model = None
                    if not self.target_printer_id and self.model:
                        target_model = VIRTUAL_PRINTER_MODELS.get(self.model)
                    # #1733: multi-plate "Send All" uploads ship every plate in
                    # one 3MF — `slice_info.config` lists each `<plate>` with
                    # its own index. Enqueue one PrintQueueItem per plate so
                    # the scheduler runs each separately. Single-plate "Send"
                    # comes through as `[N]` (one plate index) so the loop
                    # below runs once and the existing behaviour is preserved.
                    plate_ids = self._extract_plate_ids(file_path)

                    # Pick a base position the same way the manual /print-queue/
                    # POST does, then hand consecutive positions to each plate
                    # so a Send All keeps plate-order execution inside the
                    # queue (#1733). Previously hardcoded to 1, which created
                    # duplicate position=1 rows on every VP upload and made
                    # queue execution order non-deterministic for any non-
                    # empty queue.
                    from sqlalchemy import func, select as _sql_select

                    # One sequence across all pending items, not one per
                    # printer (#3200): the plates go to the end of the queue.
                    queue_scope = _sql_select(func.max(PrintQueueItem.position)).where(
                        PrintQueueItem.status == "pending"
                    )
                    try:
                        max_pos_raw = (await db.execute(queue_scope)).scalar()
                        max_pos = int(max_pos_raw) if max_pos_raw is not None else 0
                    except (TypeError, ValueError):
                        max_pos = 0

                    # Parse per-plate filament requirements (#1188). Each plate
                    # has its own filament set in `slice_info.config`, so the
                    # `required_filament_types` / `filament_overrides` columns
                    # on each queue item reflect THAT plate, not the file's
                    # first plate. Scoping was already plate-aware via #1697 —
                    # the `extract_filament_requirements(path, plate_id)` filter
                    # returns just the plate's filaments. required_filament_types
                    # is populated unconditionally — it's cheap, lets the
                    # scheduler reject obvious mis-matches even without
                    # force_color_match. filament_overrides only carries
                    # force_color_match=True when the per-VP setting is on, so
                    # upgraders keep the old behaviour by default.
                    queue_item_ids: list[int] = []
                    queue_item_reach: dict[int, int] = {}
                    for offset, plate_id in enumerate(plate_ids, start=1):
                        required_filament_types_json: str | None = None
                        filament_overrides_json: str | None = None
                        requirements = extract_filament_requirements(file_path, plate_id)
                        if requirements:
                            types = sorted({r["type"] for r in requirements if r.get("type")})
                            if types:
                                required_filament_types_json = json.dumps(types)
                            if self.queue_force_color_match:
                                # Carry tray_info_idx so force_color_match can
                                # tell Bambu PLA variants apart (#2650). Bambu
                                # reports Basic/Matte/Silk all as tray_type
                                # "PLA"; the variant lives only in tray_info_idx
                                # (GFA00/GFA01/GFA06/...). A blank idx (custom or
                                # third-party spool) means "no variant
                                # constraint" and the scheduler falls back to
                                # type+colour.
                                overrides = [
                                    {
                                        "slot_id": r["slot_id"],
                                        "type": r.get("type", ""),
                                        "color": r.get("color", ""),
                                        "tray_info_idx": r.get("tray_info_idx", ""),
                                        "force_color_match": True,
                                    }
                                    for r in requirements
                                    if r.get("type") and r.get("color")
                                ]
                                if overrides:
                                    filament_overrides_json = json.dumps(overrides)

                        # The slicer's mapping is indexed by the 3MF's own
                        # file-global slot ids (position = slot_id - 1), so one
                        # array covers every plate of a multi-plate Send All —
                        # each plate just reads the entries for the slots it
                        # actually prints. What must be checked is that it
                        # reaches that far: a mapping shorter than this plate's
                        # highest slot id can't address the plate's own slots,
                        # and `_ensure_ams_mapping` would keep it anyway
                        # because it only rejects an all-unresolved mapping. Fall
                        # back to a computed mapping for that plate instead
                        # (#2700 review).
                        plate_ams_mapping_json = queue_ams_mapping_json
                        plate_max_slot = max((r.get("slot_id") or 0) for r in requirements) if requirements else 0
                        if queue_ams_mapping is not None and requirements:
                            max_slot_id = max((r.get("slot_id") or 0) for r in requirements)
                            if max_slot_id > len(queue_ams_mapping):
                                logger.warning(
                                    "[VP %s] Slicer ams_mapping has %d entries but plate %s needs slot %d; "
                                    "dropping it for this plate so the scheduler computes one from live AMS state.",
                                    self.name,
                                    len(queue_ams_mapping),
                                    plate_id,
                                    max_slot_id,
                                )
                                plate_ams_mapping_json = None

                        queue_item = PrintQueueItem(
                            printer_id=self.target_printer_id,
                            target_model=target_model,
                            archive_id=archive.id,
                            plate_id=plate_id,
                            position=max_pos + offset,
                            status="pending",
                            manual_start=not self.auto_dispatch,
                            required_filament_types=required_filament_types_json,
                            filament_overrides=filament_overrides_json,
                            bed_levelling=bed_levelling,
                            flow_cali=flow_cali,
                            nozzle_offset_cali=nozzle_offset_cali,
                            vibration_cali=vibration_cali,
                            layer_inspect=layer_inspect,
                            timelapse=timelapse,
                            use_ams=use_ams,
                            confirm_outcome=confirm_outcome,
                            # Per-VP opt-in for auto-print G-code injection (#1516).
                            # Default off; when on, the scheduler still no-ops unless
                            # gcode_snippets are configured for the target model, so it's
                            # effectively "inject when enabled AND snippets exist".
                            gcode_injection=self.gcode_injection,
                            # H2C rack-swap slicer pick (#1780). Captured above;
                            # stamped on every plate so a multi-plate Send All keeps
                            # the same nozzle pick across plates rather than only the
                            # first one (mirrors the #1697 / #1188 per-plate loop fix).
                            nozzle_mapping=nozzle_mapping_json,
                            # Slicer's own live AMS-slot pick, when present —
                            # see `_extract_slicer_ams_mapping_json`.
                            ams_mapping=plate_ams_mapping_json,
                        )
                        db.add(queue_item)
                        await db.flush()  # populate queue_item.id before logging
                        queue_item_ids.append(queue_item.id)
                        queue_item_reach[queue_item.id] = plate_max_slot
                    await db.commit()
                    # Track the freshly-committed queue items so
                    # `on_print_command` can retroactively stamp slicer-side
                    # fields if the MQTT `project_file` lands AFTER the
                    # `_SLICER_OPTIONS_WAIT_TIMEOUT` window expired — the
                    # #1780 round-3 race. Eviction of stale entries here
                    # keeps the dict bounded; the queue path is the only
                    # writer, so doing it on commit is enough.
                    now = time.monotonic()
                    cutoff = now - _RECENT_QUEUE_ITEM_TTL
                    self._recent_queue_items = {k: v for k, v in self._recent_queue_items.items() if v[1] > cutoff}
                    self._recent_queue_items[file_path.name] = (list(queue_item_ids), now)
                    self._recent_queue_item_reach = {
                        k: v for k, v in self._recent_queue_item_reach.items() if k in self._recent_queue_items
                    }
                    self._recent_queue_item_reach[file_path.name] = queue_item_reach
                    # Last-chance check: MQTT for this filename could have
                    # arrived during ANY await between the initial pop and
                    # now — wait_for itself, archive_print, db.flush,
                    # db.commit. In all those cases `on_print_command`
                    # stashed its data but neither the event-signal path nor
                    # the retroactive `_recent_queue_items` path was in
                    # place to consume it. Pop any late stash and apply
                    # inline so the late MQTT never leaks past the queue-add.
                    late_opts = self._slicer_print_options.pop(file_path.name, None)
                    if late_opts is not None:
                        logger.info(
                            "[VP %s] Late slicer MQTT detected for %s during queue-add — "
                            "applying inline (race vs commit/archive/flush yield)",
                            self.name,
                            file_path.name,
                        )
                        await self._restamp_recent_queue_item(file_path.name, late_opts)
                    if len(queue_item_ids) == 1:
                        logger.info("[VP %s] Added to queue: %s", self.name, queue_item_ids[0])
                    else:
                        logger.info(
                            "[VP %s] Added %d queue items for multi-plate upload (plates %s): %s",
                            self.name,
                            len(queue_item_ids),
                            plate_ids,
                            queue_item_ids,
                        )
                    await self._broadcast_archive_created(archive)
                    from backend.app.services.queue_announcements import announce_queue_jobs_added

                    await announce_queue_jobs_added(db, queue_item_ids)
                else:
                    logger.error("Failed to archive file: %s", file_path.name)
        except Exception as e:
            logger.error("Error adding to print queue: %s", e)
        finally:
            # Always release the marker and clean the temp file. Without this
            # the same-name STOR guard would block the next upload and the
            # upload_dir would accumulate failed temp files forever
            # (#audit-R2-1).
            self._pending_files.pop(file_path.name, None)
            try:
                file_path.unlink(missing_ok=True)
            except OSError:
                pass

    async def _broadcast_archive_created(self, archive) -> None:
        """Notify connected clients that a new archive exists.

        Real-printer prints get this from main.py's MQTT print_start handler;
        VP-uploaded prints need their own broadcast or the Archives page stays
        stale until the user switches tabs (#1282).
        """
        try:
            from backend.app.core.websocket import ws_manager

            await ws_manager.send_archive_created(
                {
                    "id": archive.id,
                    "printer_id": archive.printer_id,
                    "filename": archive.filename,
                    "print_name": archive.print_name,
                    "status": archive.status,
                }
            )
        except Exception as e:
            logger.debug("[VP %s] archive_created broadcast failed: %s", self.name, e)

    @staticmethod
    def _extract_plate_ids(file_path: Path) -> list[int]:
        """Extract every plate index from a 3MF's slice_info.config.

        A multi-plate "Send All" from BambuStudio / OrcaSlicer uploads a
        single 3MF containing every plate the user selected. Each plate
        has its own ``<plate>`` block with a ``<metadata key="index"
        value="N"/>`` child and its own ``Metadata/plate_N.gcode`` payload
        inside the same zip. Returning the full ordered list lets the VP
        queue path create one queue item per plate (`_add_to_print_queue`
        loops over the result), so "Send All" of a 3-plate file produces
        3 queue items sharing the same archive — one per plate to print.

        Single-plate "Send" hits the same code path and returns ``[N]``
        for whichever plate the user selected; the loop runs once and the
        existing single-plate behaviour is preserved.

        Returns ``[1]`` when the 3MF is missing ``slice_info.config``,
        unparseable, or contains no plate-index metadata — the original
        single-plate fallback. Production logs at debug so a non-3MF
        upload doesn't spam, but the trail survives for support bundles.
        """
        try:
            import xml.etree.ElementTree as ET
            import zipfile

            with zipfile.ZipFile(file_path, "r") as zf:
                if "Metadata/slice_info.config" in zf.namelist():
                    content = zf.read("Metadata/slice_info.config").decode()
                    root = ET.fromstring(content)  # noqa: S314  # nosec B314
                    plate_ids: list[int] = []
                    for plate in root.findall(".//plate"):
                        for meta in plate.findall("metadata"):
                            if meta.get("key") == "index" and meta.get("value"):
                                try:
                                    plate_ids.append(int(meta.get("value")))
                                except ValueError:
                                    continue
                                break
                    if plate_ids:
                        return plate_ids
        except Exception as e:
            logger.debug("[VP] _extract_plate_ids failed for %s: %s", file_path.name, e)
        return [1]

    # -- Service lifecycle --

    def _resolve_cert_and_advertise(self) -> tuple[Path, Path, str]:
        """Return (cert_path, key_path, advertise_address) for TLS services.

        Always uses the self-signed cert chain (signed by `bbl_ca`). The user
        imports `bbl_ca.crt` once into the slicer; per-VP certs validate from
        there. Tailscale exposure is handled by the user picking the Tailscale
        IP in the bind_ip dropdown.
        """
        cert_path, key_path = self.generate_certificates()
        advertise = self.remote_interface_ip or self.bind_ip or ""
        return cert_path, key_path, advertise

    async def start_server(self) -> None:
        """Start server-mode services (FTP, MQTT, SSDP, Bind) on this VP's bind_ip."""
        logger.info("[VP %s] Starting server-mode services on %s", self.name, self.bind_ip)

        cert_path, key_path, advertise_addr = self._resolve_cert_and_advertise()
        bind_addr = self.bind_ip or "0.0.0.0"  # nosec B104

        async def run_with_logging(coro, svc_name):
            try:
                await coro
            except Exception as e:
                logger.error("[VP %s] %s failed: %s", self.name, svc_name, e)

        self._tasks = []

        # FTP server. Each VP gets a non-overlapping passive-mode port slice
        # derived from its DB id so bridge-mode Docker users only have to
        # expose a narrow range (#1646). Default slice is 10 ports per VP;
        # see ftp_server.compute_passive_port_slice for the wrap-around
        # behaviour on installs with very high VP ids.
        passive_port_min, passive_port_max = compute_passive_port_slice(self.id)
        self._ftp = VirtualPrinterFTPServer(
            upload_dir=self.upload_dir,
            access_code=self.access_code,
            cert_path=cert_path,
            key_path=key_path,
            on_file_received=self.on_file_received,
            bind_address=bind_addr,
            vp_name=self.name,
            passive_port_min=passive_port_min,
            passive_port_max=passive_port_max,
        )
        self._tasks.append(
            asyncio.create_task(
                run_with_logging(self._ftp.start(), "FTP"),
                name=f"vp_{self.id}_ftp",
            )
        )

        # MQTT server
        self._mqtt = SimpleMQTTServer(
            serial=self.serial,
            access_code=self.access_code,
            cert_path=cert_path,
            key_path=key_path,
            on_print_command=self.on_print_command,
            model=self.model or DEFAULT_VIRTUAL_PRINTER_MODEL,
            bind_address=bind_addr,
            vp_name=self.name,
        )
        self._tasks.append(
            asyncio.create_task(
                run_with_logging(self._mqtt.start(), "MQTT"),
                name=f"vp_{self.id}_mqtt",
            )
        )

        # MQTT bridge — fans out the target printer's pushes to slicers connected
        # to this VP and forwards their commands back to the printer. Only meaningful
        # when a target printer is configured AND printer_manager was injected (it
        # always is at runtime; tests may omit it).
        if self.target_printer_id is not None and self._printer_manager is not None:
            self._mqtt_bridge = MQTTBridge(
                vp_id=self.id,
                vp_name=self.name,
                vp_serial=self.serial,
                target_printer_id=self.target_printer_id,
                mqtt_server=self._mqtt,
                printer_manager=self._printer_manager,
            )
            self._mqtt.set_bridge(self._mqtt_bridge)
            await self._mqtt_bridge.start()

            # Camera passthrough. BambuStudio / OrcaSlicer connect the "camera"
            # button to the device IP they bound on (the VP), not the IP in the
            # printer's `ipcam.rtsp_url`. Without a listener the slicer gets
            # connection refused → "LAN connection failed" (RTSP models) or
            # OrcaSlicer error `[2:-10061]` (chamber-image models, #1868).
            #
            # The port depends on the TARGET printer's model:
            #   RTSPS (X1/X2/H2/P2S)        → 322
            #   chamber-image (A1/P1P/P1S)  → 6000
            #
            # `get_camera_port()` is the same source of truth used by
            # `routes/camera.py`, so slicer and Bambuddy UI agree.
            target_client = self._printer_manager.get_client(self.target_printer_id)
            target_ip = getattr(target_client, "ip_address", None) if target_client else None
            target_model = getattr(target_client, "model", None) if target_client else None
            if target_ip:
                from backend.app.services.camera import get_camera_port

                camera_port = get_camera_port(target_model)
                self._rtsp_proxy = TCPProxy(
                    name=f"Camera-{camera_port}",
                    listen_port=camera_port,
                    target_host=target_ip,
                    target_port=camera_port,
                    bind_address=bind_addr,
                )
                self._tasks.append(
                    asyncio.create_task(
                        run_with_logging(self._rtsp_proxy.start(), f"Camera-{camera_port}"),
                        name=f"vp_{self.id}_camera",
                    )
                )

        # Bind server
        self._bind = BindServer(
            serial=self.serial,
            model=self.model or DEFAULT_VIRTUAL_PRINTER_MODEL,
            name=self.name,
            bind_address=bind_addr,
            cert_path=cert_path,
            key_path=key_path,
        )
        self._tasks.append(
            asyncio.create_task(
                run_with_logging(self._bind.start(), "Bind"),
                name=f"vp_{self.id}_bind",
            )
        )

        # SSDP server — advertise_addr is the remote_interface_ip (Tailscale
        # IP, when chosen from the bind_ip dropdown) or the bind_ip. SSDP
        # Location accepts IPs only; FQDNs go in through bind_ip selection
        # at the printer-IP level and resolve before reaching the SSDP
        # advertisement.
        self._ssdp = VirtualPrinterSSDPServer(
            name=self.name,
            serial=self.serial,
            model=self.model or DEFAULT_VIRTUAL_PRINTER_MODEL,
            advertise_ip=advertise_addr,
            bind_ip=bind_addr,
        )
        self._tasks.append(
            asyncio.create_task(
                run_with_logging(self._ssdp.start(), "SSDP"),
                name=f"vp_{self.id}_ssdp",
            )
        )

        # Wait briefly for every child service to actually finish binding its
        # socket so ``is_running`` doesn't lie. Without this barrier a caller
        # racing the start (e.g. the diagnostic route) would see is_running=True
        # while ports were still in the gap between task creation and the
        # ``asyncio.start_server`` returning. Bounded timeout — if a child
        # hangs we log it and move on; the existing task tracking still
        # catches the failure on the next iteration.
        ready_targets = [
            ("FTP", self._ftp.ready),
            ("MQTT", self._mqtt.ready),
            ("Bind", self._bind.ready),
            ("SSDP", self._ssdp.ready),
        ]
        try:
            await asyncio.wait_for(
                asyncio.gather(*(e.wait() for _, e in ready_targets)),
                timeout=5.0,
            )
        except TimeoutError:
            not_ready = [name for name, e in ready_targets if not e.is_set()]
            logger.warning(
                "[VP %s] Sub-service(s) didn't bind within 5s: %s — continuing anyway",
                self.name,
                ", ".join(not_ready) or "(none)",
            )

        logger.info("[VP %s] Server-mode services started on %s", self.name, bind_addr)

    async def stop_server(self) -> None:
        """Stop server-mode services."""
        if self._finish_release_task is not None and not self._finish_release_task.done():
            self._finish_release_task.cancel()
            self._finish_release_task = None
        if self._mqtt_bridge:
            try:
                await self._mqtt_bridge.stop()
            except Exception:
                logger.exception("[VP %s] MQTT bridge stop failed", self.name)
            if self._mqtt:
                self._mqtt.set_bridge(None)
            self._mqtt_bridge = None
        if self._rtsp_proxy:
            try:
                await self._rtsp_proxy.stop()
            except Exception:
                logger.exception("[VP %s] Camera proxy stop failed", self.name)
            self._rtsp_proxy = None
        if self._ftp:
            await self._ftp.stop()
            self._ftp = None
        if self._mqtt:
            await self._mqtt.stop()
            self._mqtt = None
        if self._bind:
            await self._bind.stop()
            self._bind = None
        if self._ssdp:
            await self._ssdp.stop()
            self._ssdp = None
        await self._cancel_tasks()

    async def start_proxy(self) -> None:
        """Start proxy mode services for this instance."""
        logger.info("[VP %s] Starting proxy mode to %s", self.name, self.target_printer_ip)

        cert_path, key_path, _ = self._resolve_cert_and_advertise()

        self._proxy = SlicerProxyManager(
            target_host=self.target_printer_ip,
            cert_path=cert_path,
            key_path=key_path,
            on_activity=lambda n, m: logger.info("[VP %s] Proxy %s: %s", self.name, n, m),
            bind_address=self.bind_ip or "0.0.0.0",  # nosec B104
            bind_identity={
                "serial": self.target_printer_serial or self.serial,
                "model": self.model or DEFAULT_VIRTUAL_PRINTER_MODEL,
                "name": self.name,
                "version": "01.00.00.00",
            },
        )

        async def run_with_logging(coro, svc_name):
            try:
                await coro
            except Exception as e:
                logger.error("[VP %s] %s failed: %s", self.name, svc_name, e)

        self._tasks = []

        # SSDP for proxy
        proxy_serial = self.target_printer_serial or self.serial
        if self.remote_interface_ip:
            from backend.app.services.network_utils import find_interface_for_ip

            local_iface = find_interface_for_ip(self.target_printer_ip)
            if local_iface:
                self._ssdp_proxy = SSDPProxy(
                    local_interface_ip=local_iface["ip"],
                    remote_interface_ip=self.remote_interface_ip,
                    target_printer_ip=self.target_printer_ip,
                    name=self.name,
                )
                self._tasks.append(
                    asyncio.create_task(
                        run_with_logging(self._ssdp_proxy.start(), "SSDP Proxy"),
                        name=f"vp_{self.id}_ssdp_proxy",
                    )
                )
            else:
                self._start_fallback_ssdp(proxy_serial, run_with_logging)
        else:
            self._start_fallback_ssdp(proxy_serial, run_with_logging)

        self._tasks.append(
            asyncio.create_task(
                run_with_logging(self._proxy.start(), "Proxy"),
                name=f"vp_{self.id}_proxy",
            )
        )

    def _start_fallback_ssdp(self, proxy_serial: str, run_with_logging) -> None:
        """Start single-interface SSDP server as fallback for proxy mode."""
        self._ssdp = VirtualPrinterSSDPServer(
            name=f"{self.name} (Proxy)",
            serial=proxy_serial,
            model=self.model or DEFAULT_VIRTUAL_PRINTER_MODEL,
            advertise_ip=self.bind_ip or "",
            bind_ip=self.bind_ip or "",
        )
        self._tasks.append(
            asyncio.create_task(
                run_with_logging(self._ssdp.start(), "SSDP"),
                name=f"vp_{self.id}_ssdp",
            )
        )

    async def stop_proxy(self) -> None:
        """Stop proxy mode services for this instance."""
        if self._proxy:
            await self._proxy.stop()
            self._proxy = None
        if self._ssdp:
            await self._ssdp.stop()
            self._ssdp = None
        if self._ssdp_proxy:
            await self._ssdp_proxy.stop()
            self._ssdp_proxy = None
        await self._cancel_tasks()

    async def _cancel_tasks(self) -> None:
        """Cancel all running tasks and wait for cleanup."""
        for task in self._tasks:
            task.cancel()
        if self._tasks:
            try:
                await asyncio.wait_for(asyncio.gather(*self._tasks, return_exceptions=True), timeout=1.0)
            except TimeoutError:
                pass
        self._tasks = []

    def get_status(self) -> dict:
        """Get status for this instance."""
        status: dict = {
            "running": self.is_running,
            "pending_files": len(self._pending_files),
        }
        if self.is_proxy and self._proxy:
            status["proxy"] = self._proxy.get_status()
        return status


class VirtualPrinterManager:
    """Multi-instance virtual printer registry and orchestrator.

    Every VP runs its own independent services on a dedicated bind IP.
    """

    def __init__(self):
        self._session_factory: Callable | None = None
        self._printer_manager: PrinterManager | None = None
        self._instances: dict[int, VirtualPrinterInstance] = {}
        # Serialize sync_from_db so concurrent PUT /vp/{id} calls can't
        # race the start/stop sequence and leave duplicate sub-services
        # bound to the same port. The lock is fine-grained enough that
        # a single VP update completes in well under a second; if the
        # user holds the lock with a long-running start they intended
        # to anyway.
        self._sync_lock = asyncio.Lock()

        # Directories
        self._base_dir = app_settings.base_dir / "virtual_printer"

        # Ensure base directories exist
        self._ensure_base_directories()

    def _ensure_base_directories(self) -> None:
        """Create base directories at startup."""
        for dir_path in [self._base_dir, self._base_dir / "uploads", self._base_dir / "certs"]:
            try:
                dir_path.mkdir(parents=True, exist_ok=True)
            except PermissionError:
                logger.error(
                    f"Cannot create directory {dir_path}: Permission denied. "
                    f"For Docker: ensure the data volume is writable by the container user. "
                    f"For bare metal: run 'sudo chown -R $(whoami) {self._base_dir}'"
                )

    def set_session_factory(self, session_factory: Callable) -> None:
        """Set the database session factory."""
        self._session_factory = session_factory

    def set_printer_manager(self, printer_manager: "PrinterManager") -> None:
        """Inject the global printer_manager so non-proxy VPs can mirror their target's MQTT stream."""
        self._printer_manager = printer_manager

    def get_ca_certificate_info(self) -> dict:
        """Return the shared virtual-printer CA certificate for slicer-trust import.

        The CA is shared by every VP (one import covers all of them). It is
        generated on demand here if no VP has triggered cert generation yet,
        so the "copy/download certificate" UI works even before the first VP
        is enabled.
        """
        certs_dir = self._base_dir / "certs"
        cert_service = CertificateService(cert_dir=certs_dir, shared_ca_dir=certs_dir)
        return cert_service.get_ca_certificate_info()

    @property
    def is_enabled(self) -> bool:
        """Check if any virtual printer is running."""
        return len(self._instances) > 0

    async def sync_from_db(self) -> None:
        """Load all VPs from DB, reconcile running state.

        Serialised by ``self._sync_lock`` — concurrent PUT /vp/{id} routes
        all call into this method; without the lock the start / stop
        sequence races and can leave duplicate sub-services bound to the
        same port or orphan still-running tasks.
        """
        if not self._session_factory:
            logger.warning("Cannot sync virtual printers: no session factory")
            return

        async with self._sync_lock:
            await self._sync_from_db_locked()

    async def _sync_from_db_locked(self) -> None:
        """Inner sync body — caller holds ``self._sync_lock``."""
        from sqlalchemy import select

        from backend.app.models.printer import Printer
        from backend.app.models.virtual_printer import VirtualPrinter

        async with self._session_factory() as db:
            result = await db.execute(
                select(VirtualPrinter).where(VirtualPrinter.enabled == True).order_by(VirtualPrinter.position)  # noqa: E712
            )
            enabled_vps = result.scalars().all()

        # Stop instances that are no longer enabled or changed mode
        enabled_ids = {vp.id for vp in enabled_vps}
        for vp_id in list(self._instances.keys()):
            if vp_id not in enabled_ids:
                await self.remove_instance(vp_id)

        # Look up printer IPs for proxy VPs
        proxy_vps = [vp for vp in enabled_vps if vp.mode == "proxy"]
        proxy_ips: dict[int, tuple[str, str]] = {}
        if proxy_vps:
            async with self._session_factory() as db:
                for pvp in proxy_vps:
                    if pvp.target_printer_id:
                        result = await db.execute(select(Printer).where(Printer.id == pvp.target_printer_id))
                        printer = result.scalar_one_or_none()
                        if printer:
                            proxy_ips[pvp.id] = (printer.ip_address, printer.serial_number)

        # Detect config changes on running instances and restart if needed
        for vp in enabled_vps:
            instance = self._instances.get(vp.id)
            if not instance:
                continue

            # Proxy mode: detect target printer IP / serial changes from the
            # DB lookup above. Without this branch a DHCP renewal that gives
            # the target printer a new IP would leave the running proxy
            # forwarding to the stale IP until the user manually toggles the
            # VP. The same shape covers a target-side serial change.
            proxy_target_changed = False
            if vp.mode == "proxy":
                fresh = proxy_ips.get(vp.id)
                if fresh is not None:
                    fresh_ip, fresh_serial = fresh
                    if (
                        getattr(instance, "target_printer_ip", None) != fresh_ip
                        or getattr(instance, "target_printer_serial", None) != fresh_serial
                    ):
                        proxy_target_changed = True

            # Normalize the DB value before comparing — a legacy `immediate`
            # row read before the migration window finishes would otherwise
            # trip the "changed" branch and bounce every VP at boot.
            db_mode = normalize_vp_mode(vp.mode)
            changed = (
                instance.mode != db_mode
                # The name is advertised to the slicer (SSDP, bind) by the
                # running services, so a rename only reaches it on restart.
                or instance.name != vp.name
                or instance.model != (vp.model or DEFAULT_VIRTUAL_PRINTER_MODEL)
                or instance.access_code != (vp.access_code or "")
                or instance.bind_ip != (vp.bind_ip or "")
                or instance.remote_interface_ip != (vp.remote_interface_ip or "")
                or instance.target_printer_id != vp.target_printer_id
                or instance.auto_dispatch != vp.auto_dispatch
                # Queue-mode behaviour toggle — without it the running
                # instance silently keeps the old value until process
                # restart (#1552 follow-up family).
                or instance.queue_force_color_match != vp.queue_force_color_match
                or instance.save_ams_mapping != vp.save_ams_mapping
                or instance.gcode_injection != vp.gcode_injection
                or proxy_target_changed
            )

            if changed:
                logger.info(
                    "VP %s config changed (mode: %s→%s), restarting",
                    instance.name,
                    instance.mode,
                    vp.mode,
                )
                await self.remove_instance(vp.id)

        # Start instances for all enabled VPs (skip already running)
        for vp in enabled_vps:
            if vp.id in self._instances:
                continue

            if vp.mode == "proxy":
                ip_info = proxy_ips.get(vp.id)
                if not ip_info:
                    logger.warning("Proxy VP %s: target printer not found, skipping", vp.name)
                    continue
                target_ip, target_serial = ip_info
                instance = VirtualPrinterInstance(
                    vp_id=vp.id,
                    name=vp.name,
                    mode=vp.mode,
                    model=vp.model or DEFAULT_VIRTUAL_PRINTER_MODEL,
                    access_code=vp.access_code or "",
                    serial_suffix=vp.serial_suffix,
                    target_printer_ip=target_ip,
                    target_printer_serial=target_serial,
                    # Carried although the proxy does not use them: the change
                    # check above compares them with the row, and a missing
                    # value restarted every proxy VP on every sync.
                    target_printer_id=vp.target_printer_id,
                    auto_dispatch=vp.auto_dispatch,
                    queue_force_color_match=vp.queue_force_color_match,
                    save_ams_mapping=vp.save_ams_mapping,
                    gcode_injection=vp.gcode_injection,
                    bind_ip=vp.bind_ip or "",
                    remote_interface_ip=vp.remote_interface_ip or "",
                    tailscale_disabled=vp.tailscale_disabled,
                    base_dir=self._base_dir,
                    session_factory=self._session_factory,
                )
                self._instances[vp.id] = instance
                await instance.start_proxy()
                logger.info("Started proxy VP: %s → %s (bind=%s)", instance.name, target_ip, instance.bind_ip)
            else:
                instance = VirtualPrinterInstance(
                    vp_id=vp.id,
                    name=vp.name,
                    mode=vp.mode,
                    model=vp.model or DEFAULT_VIRTUAL_PRINTER_MODEL,
                    access_code=vp.access_code or "",
                    serial_suffix=vp.serial_suffix,
                    target_printer_id=vp.target_printer_id,
                    auto_dispatch=vp.auto_dispatch,
                    queue_force_color_match=vp.queue_force_color_match,
                    save_ams_mapping=vp.save_ams_mapping,
                    gcode_injection=vp.gcode_injection,
                    bind_ip=vp.bind_ip or "",
                    remote_interface_ip=vp.remote_interface_ip or "",
                    tailscale_disabled=vp.tailscale_disabled,
                    base_dir=self._base_dir,
                    session_factory=self._session_factory,
                    printer_manager=self._printer_manager,
                )
                self._instances[vp.id] = instance
                await instance.start_server()
                logger.info("Started server-mode VP: %s on %s", instance.name, vp.bind_ip)

    async def remove_instance(self, vp_id: int) -> None:
        """Stop and remove a single VP instance."""
        instance = self._instances.pop(vp_id, None)
        if instance:
            if instance.is_proxy:
                await instance.stop_proxy()
            else:
                await instance.stop_server()
            logger.info("Removed VP instance: %s", instance.name)

    async def stop_all(self) -> None:
        """Shutdown all virtual printer services."""
        logger.info("Stopping all virtual printer services...")

        for vp_id in list(self._instances.keys()):
            await self.remove_instance(vp_id)

        logger.info("All virtual printer services stopped")

    def get_instance(self, vp_id: int) -> VirtualPrinterInstance | None:
        """Get a running instance by ID."""
        return self._instances.get(vp_id)

    def get_all_status(self) -> list[dict]:
        """Get status for all running instances."""
        return [
            {
                "id": inst.id,
                "name": inst.name,
                "mode": inst.mode,
                **inst.get_status(),
            }
            for inst in self._instances.values()
        ]

    # -- Legacy single-printer compat --

    def get_status(self) -> dict:
        """Get status for first virtual printer (backward compat)."""
        if self._instances:
            first = next(iter(self._instances.values()))
            return {
                "enabled": True,
                "running": first.is_running,
                "mode": first.mode,
                "name": first.name,
                "serial": first.serial,
                "model": first.model or DEFAULT_VIRTUAL_PRINTER_MODEL,
                "model_name": VIRTUAL_PRINTER_MODELS.get(
                    first.model or DEFAULT_VIRTUAL_PRINTER_MODEL,
                    first.model or DEFAULT_VIRTUAL_PRINTER_MODEL,
                ),
                "pending_files": first.get_status().get("pending_files", 0),
                **({"target_printer_ip": first.target_printer_ip} if first.is_proxy else {}),
                **({"proxy": first.get_status().get("proxy", {})} if first.is_proxy else {}),
            }
        return {
            "enabled": False,
            "running": False,
            "mode": VP_MODE_ARCHIVE,
            "name": "Bambuddy",
            "serial": "",
            "model": DEFAULT_VIRTUAL_PRINTER_MODEL,
            "model_name": VIRTUAL_PRINTER_MODELS[DEFAULT_VIRTUAL_PRINTER_MODEL],
            "pending_files": 0,
        }

    async def configure(
        self,
        enabled: bool,
        access_code: str = "",
        mode: str = VP_MODE_ARCHIVE,
        model: str = "",
        target_printer_ip: str = "",
        target_printer_serial: str = "",
        remote_interface_ip: str = "",
    ) -> None:
        """Legacy single-printer configure. Delegates to sync_from_db()."""
        # This method is kept for backward compat with the settings endpoint.
        # The actual work is done by sync_from_db() which reads from the DB.
        await self.sync_from_db()


# Global instance
virtual_printer_manager = VirtualPrinterManager()