fix(deploy): eGalax touchscreen on batm3 (kernel 6.6 + calibration)

The Dell 9030 AIO's built-in eGalax SAW panel (0eef:0001) was unusable:
touches either didn't register or landed in the wrong place. Full fix:

- Pin linuxPackages_6_6. On 25.11's default 6.12 kernel hid-multitouch
  grabs the controller and mis-parses its HID report (axes read stuck)
  and usbtouchscreen refuses to bind. On 6.6 usbtouchscreen binds and
  produces a clean single-touch ABS device (the known-good internal-SATA
  install runs 6.6.68). Mirrors douro.nix's per-hardware kernel pin.

- udev rule now modprobes usbtouchscreen ITSELF before unbinding usbhid
  and handing over via new_id. On a USB boot systemd-udev-trigger fires
  this rule (~2s) before systemd-modules-load loads usbtouchscreen
  (~12s), so new_id previously hit a not-yet-loaded driver and the panel
  bound to nothing. Loading it inline removes the boot-ordering race.

- Add an X evdev InputClass (99-egalax.conf) so X uses evdev + the
  transformation matrix rather than libinput. Mirrors the working
  internal-SATA install.

- egalax-calibrate: add XAUTHORITY (=/home/bitspire/.Xauthority) — the
  actual boot-time bug. Without the auth cookie xinput died with
  "Invalid MIT-MAGIC-COOKIE-1 key / Unable to connect to X server", so
  the coordinate-transformation matrix was never applied and touches
  landed in the wrong place. Also replace the fixed ExecStartPre sleep
  with a 30s retry loop on the eGalax X device appearing — more robust
  to boot timing than a race against display-manager.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

apps/machine/electron/hal-service.ts

@@ -221,6 +221,15 @@ export async function initializeHal(config: HalConfig): Promise<HalInstance> {
     },
 
     disableValidator: () => {
+      // If a note is sitting in escrow when we disable (inactivity timeout,
+      // cancel, or leaving the insert screen), return it to the customer.
+      // Disabling alone does NOT release an escrowed note on EBDS — it would
+      // be stranded in the transport until the next power cycle.
+      if (escrowDenomination !== null) {
+        console.log('[HAL] Returning escrowed bill on disable:', escrowDenomination)
+        escrowDenomination = null
+        validator?.reject()
+      }
       validator?.disable()
       validator?.lightOff()
     },

apps/machine/src/config/device.ts

@@ -139,11 +139,21 @@ export const MACHINE_PRESETS: Record<MachineModel, Omit<DeviceConfig, 'fiatCode'
     model: 'batm3',
     validator: {
       type: 'ebds',
-      device: '/dev/ttyACM0',
+      // MEI bill acceptor (EBDS) on a USB-serial bridge, exposed via the
+      // stable udev symlink /dev/ttyMEI (batm3.nix, serial A9YW78OC). The old
+      // /dev/ttyACM0 default assumed a CDC-ACM BNR; this hardware enumerates as
+      // ttyUSB* instead, so ACM0 never existed and cash-in was silently
+      // disabled ("[HAL] No validator"). Per-box override: VITE_LAMASSU_VALIDATOR_DEVICE.
+      device: '/dev/ttyMEI',
     },
     dispenser: {
       type: 'f56',
-      device: '/dev/ttyUSB0',
+      // Fujitsu F56 on a USB-serial bridge, via the stable udev symlink
+      // /dev/ttyF56 (batm3.nix, serial DDDLb103Y23). Avoids the raw
+      // /dev/ttyUSB0, which is enumeration-order dependent and could point at
+      // the wrong adapter after a re-plug/reboot. Per-box override:
+      // VITE_LAMASSU_DISPENSER_DEVICE.
+      device: '/dev/ttyF56',
       cassettes: [
         { denomination: 20, count: 400 },
         { denomination: 1, count: 400 },

apps/machine/src/stores/atm.ts

@@ -458,7 +458,7 @@ export const useAtmStore = defineStore('atm', () => {
     actor.value.subscribe((newSnapshot: SnapshotFrom<ATMMachine>) => {
       const prevSnapshot = snapshot.value
       snapshot.value = newSnapshot
-      console.log('[ATM] State:', newSnapshot.value)
+      console.log('[ATM] State:', JSON.stringify(newSnapshot.value))
 
       // Detect transition into a complete state
       const state = newSnapshot.value
@@ -1374,7 +1374,7 @@ export const useAtmStore = defineStore('atm', () => {
       console.error('[ATM] Cannot send event: machine not initialized')
       return
     }
-    console.log('[ATM] Sending event:', event)
+    console.log('[ATM] Sending event:', event.type, JSON.stringify(event))
     actor.value.send(event)
   }
 

deploy/nixos/hardware/batm3.nix

@@ -12,13 +12,36 @@
       timeout = 3;
     };
 
+    # Pin the 6.6 LTS kernel. The Dell 9030 AIO's eGalax SAW touch panel
+    # (0eef:0001) works with the usbtouchscreen driver on 6.6 (the known-good
+    # internal-SATA install runs 6.6.68). On 25.11's default 6.12 kernel this
+    # old controller regressed: hid-multitouch grabs it and mis-parses the HID
+    # report ("failed to fetch feature 7", axes read stuck), usbtouchscreen
+    # refuses it, and touch is unusable regardless of udev/X config. Matching
+    # douro.nix's per-hardware kernel pin. Re-test touch before bumping this.
+    kernelPackages = pkgs.linuxPackages_6_6;
+
     initrd.availableKernelModules = [
       "xhci_pci"
       "ahci"
       "usbhid"
       "sd_mod"
+      # USB mass-storage: required to boot the dd'd image from a USB stick
+      # (stage-1 must bind the flash drive as a SCSI disk so
+      # /dev/disk/by-label/nixos appears). Harmless on the internal-SATA
+      # install, where ahci+sd_mod already cover the root device.
+      #
+      # NOTE: deliberately NO "uas" here. Many USB sticks/bridges advertise
+      # UAS but drop off the bus ("device offline error, dev sdb") under the
+      # sustained write load of first-boot growPartition/journal/swapfile.
+      # Blacklisting uas below forces the slower-but-reliable usb-storage
+      # (Bulk-Only Transport) path. SATA/eMMC installs don't use uas anyway.
+      "usb_storage"
     ];
 
+    # Keep the USB flash drive off the flaky UAS driver (see note above).
+    blacklistedKernelModules = [ "uas" ];
+
     kernelModules = [
       "kvm-intel"
       "usbtouchscreen"
@@ -27,6 +50,9 @@
     kernelParams = [
       "quiet"
       "splash"
+      # Disable USB autosuspend so the boot medium (and kiosk peripherals)
+      # aren't power-suspended mid-I/O — another cause of "device offline".
+      "usbcore.autosuspend=-1"
     ];
   };
 
@@ -105,10 +131,20 @@
   '';
 
   # eGalax touchscreen (Dell 9030 AIO built-in panel)
-  # The eGalax HID descriptor confuses libinput (treats it as touchpad).
+  # By default usbhid/hid-multitouch claim the eGalax and mis-parse its
-  # Fix: unbind from usbhid at boot, bind to usbtouchscreen kernel module,
+  # HID report descriptor (X axis reads as stuck), so touch is unusable.
-  # then apply calibration matrix after X11 starts.
+  # Fix: hand the device to the usbtouchscreen kernel driver, which parses
-  # Unbind eGalax from usbhid, bind to usbtouchscreen
+  # the raw eGalax protocol into a clean single-touch ABS device that the
+  # X evdev driver + calibration matrix (below) map correctly. This mirrors
+  # the known-good internal-SATA install.
+  #
+  # The RUN command modprobes usbtouchscreen ITSELF before unbinding usbhid
+  # and handing over via new_id. usbtouchscreen is also in boot.kernelModules
+  # (systemd-modules-load), but on a USB boot systemd-udev-trigger fires this
+  # rule (~2s) BEFORE modules-load gets usbtouchscreen in (~12s) — so the
+  # new_id write hit a not-yet-loaded driver and the panel was left bound to
+  # nothing. Loading it inline here makes the handoff independent of that
+  # boot-ordering race (on internal-SATA boot the order happened to work).
   services.udev.extraRules = lib.mkAfter ''
     KERNEL=="ttyS[0-9]*", MODE="0666"
     KERNEL=="ttyUSB[0-9]*", MODE="0666"
@@ -116,7 +152,25 @@
     SUBSYSTEM=="tty", ATTRS{serial}=="DDDLb103Y23", SYMLINK+="ttyF56", MODE="0666"
     SUBSYSTEM=="tty", ATTRS{serial}=="A9YW78OC", SYMLINK+="ttyMEI", MODE="0666"
     SUBSYSTEM=="tty", ATTRS{serial}=="A9ZF8ELY", SYMLINK+="ttyNFC", MODE="0666"
-    ACTION=="add", SUBSYSTEM=="usb", ATTRS{idVendor}=="0eef", ATTRS{idProduct}=="0001", RUN+="${pkgs.bash}/bin/bash -c 'echo ''$kernel:1.0 > /sys/bus/usb/drivers/usbhid/unbind 2>/dev/null; echo 0eef 0001 > /sys/bus/usb/drivers/usbtouchscreen/new_id 2>/dev/null'"
+    ACTION=="add", SUBSYSTEM=="usb", ATTRS{idVendor}=="0eef", ATTRS{idProduct}=="0001", RUN+="${pkgs.bash}/bin/bash -c '${pkgs.kmod}/bin/modprobe usbtouchscreen 2>/dev/null; echo ''$kernel:1.0 > /sys/bus/usb/drivers/usbhid/unbind 2>/dev/null; echo 0eef 0001 > /sys/bus/usb/drivers/usbtouchscreen/new_id 2>/dev/null'"
+  '';
+
+  # Force the X evdev driver on the eGalax (not libinput). The usbtouchscreen
+  # node is a plain single-touch absolute device; evdev + the transformation
+  # matrix in egalax-calibrate below give correct orientation. Mirrors the
+  # working internal-SATA install's /etc/X11/xorg.conf.d/99-egalax.conf.
+  environment.etc."X11/xorg.conf.d/99-egalax.conf".text = ''
+    Section "InputClass"
+        Identifier "eGalax Touchscreen"
+        MatchVendor "0eef"
+        MatchProduct "0001"
+        MatchDevicePath "/dev/input/event*"
+        Driver "evdev"
+        Option "InvertY" "false"
+        Option "InvertX" "false"
+        Option "SwapAxes" "false"
+        Option "Calibration" ""
+    EndSection
   '';
 
   # Apply touchscreen calibration after X11 starts
@@ -130,9 +184,26 @@
       Type = "oneshot";
       RemainAfterExit = true;
       User = "bitspire";
-      Environment = "DISPLAY=:0";
+      # DISPLAY *and* XAUTHORITY — without the auth cookie xinput dies with
-      ExecStartPre = "${pkgs.coreutils}/bin/sleep 3";
+      # "Invalid MIT-MAGIC-COOKIE-1 key / Unable to connect to X server" and
-      ExecStart = "${pkgs.xorg.xinput}/bin/xinput set-prop 'eGalax Inc. USB TouchController' 'Coordinate Transformation Matrix' 0 -1.268 1.147 -1.224 0 1.118 0 0 1";
+      # the matrix is never applied, so touches register but land in the wrong
+      # place (the panel then feels dead). This was the actual boot-time bug.
+      Environment = [ "DISPLAY=:0" "XAUTHORITY=/home/bitspire/.Xauthority" ];
+      # Wait for the eGalax X device to appear (usbtouchscreen binds a little
+      # after display-manager on a USB boot) and retry, instead of a fixed
+      # sleep — more robust to boot timing. Matrix: swap X/Y + invert + scale
+      # to the active panel area (matches the known-good internal install).
+      ExecStart = pkgs.writeShellScript "egalax-calibrate" ''
+        for i in $(${pkgs.coreutils}/bin/seq 1 30); do
+          if ${pkgs.xorg.xinput}/bin/xinput list --name-only 2>/dev/null | ${pkgs.gnugrep}/bin/grep -qx 'eGalax Inc. USB TouchController'; then
+            exec ${pkgs.xorg.xinput}/bin/xinput set-prop 'eGalax Inc. USB TouchController' \
+              'Coordinate Transformation Matrix' 0 -1.268 1.147 -1.224 0 1.118 0 0 1
+          fi
+          ${pkgs.coreutils}/bin/sleep 1
+        done
+        echo "egalax-calibrate: eGalax device not found after 30s" >&2
+        exit 1
+      '';
     };
   };
 

flake.nix

@@ -424,6 +424,83 @@
               printf 'verify ESP label: '; mlabel -i "$out/nixos.img@@$espStart" -s :: || true
             '';
 
+        # USB-bootable BATM3 TEST image with DISTINCT partition labels
+        # (nixos-usb / ESP-USB). The plain disk-image-batm3 reuses the generic
+        # nixos/ESP labels, so a USB stick carrying it, booted on a batm3 whose
+        # internal SATA drive ALREADY holds a nixos/ESP-labelled install, makes
+        # stage-1's by-label/nixos resolve to the internal drive (larger fs,
+        # journal recovers) instead of the stick — the stage-2 init path baked
+        # into the USB's boot entry isn't on that root, so stage 1 aborts.
+        # Distinct labels make stage-1 pick the stick unambiguously WITHOUT
+        # touching the internal drive. Unlike disk-image-sintra-usb this keeps
+        # systemd-boot: the batm3 firmware UEFI-USB-boots fine via the ESP's
+        # /EFI/BOOT/BOOTX64.EFI removable fallback, so no GRUB/hybrid-table
+        # change is needed — only the label disambiguation here plus the
+        # usb_storage/uas initrd modules (in batm3.nix). Does NOT grow to fill
+        # the stick (see the growPartition note below — sfdisk on first boot
+        # wedges flaky USB bridges); auto-upgrade off (test image, not a managed
+        # fleet member — also stops scheduled bootloader writes landing on the
+        # internal drive's ESP).
+        disk-image-batm3-usb =
+          let
+            cfg = self.nixosConfigurations.batm3-installed.extendModules {
+              modules = [
+                ({ lib, ... }: {
+                  fileSystems."/".device = lib.mkForce "/dev/disk/by-label/nixos-usb";
+                  fileSystems."/boot".device = lib.mkForce "/dev/disk/by-label/ESP-USB";
+                  # /boot must NOT be a hard boot dependency on the USB test
+                  # image. The firmware already loaded the bootloader from the
+                  # ESP before Linux started; /boot is only remounted so the OS
+                  # can *update* the bootloader — which this image never does
+                  # (autoUpgrade off, no nixos-rebuild on the stick). Without
+                  # nofail, a slow/late ESP-USB enumeration (BOT is slower than
+                  # UAS) blows past systemd's 90s device-timeout and drops to
+                  # emergency mode — with root locked, an unrecoverable dead end.
+                  # nofail + a short timeout lets the (already-mounted) root carry
+                  # the boot to completion; /boot mounts if/when the ESP shows up.
+                  fileSystems."/boot".options = [ "nofail" "x-systemd.device-timeout=10s" ];
+                  # DELIBERATELY NO growPartition/autoResize on the USB image.
+                  # growPartition runs sfdisk to rewrite the stick's partition
+                  # table on first boot — the single most bus-stressing write of
+                  # the boot. Flaky USB bridges drop off the bus mid-rewrite
+                  # (sfdisk hangs forever as an uninterruptible D-state task) and,
+                  # worse, partition 1 (ESP-USB) vanishes with the device, so
+                  # /boot times out too. The kiosk's persistent state (state.db,
+                  # .env, wifi.conf, logs) is a few MB and the built image already
+                  # carries ~2GB free inside root — growing to fill the stick buys
+                  # nothing and costs reliability. The internal-SATA target
+                  # (disk-image-batm3) keeps growPartition: a real AHCI SSD won't
+                  # drop the bus and there filling the disk is worth it.
+                  system.autoUpgrade.enable = lib.mkForce false;
+                })
+              ];
+            };
+            baseImage = import (nixpkgs + "/nixos/lib/make-disk-image.nix") {
+              inherit pkgs lib;
+              config = cfg.config;
+              format = "raw";
+              partitionTableType = "efi";
+              diskSize = "auto";
+              label = "nixos-usb"; # ext4 root label (make-disk-image -L)
+            };
+          in
+          pkgs.runCommand "nixos-disk-image-batm3-usb"
+            { nativeBuildInputs = [ pkgs.parted pkgs.mtools ]; }
+            ''
+              mkdir -p $out
+              cp --sparse=always ${baseImage}/nixos.img $out/nixos.img
+              chmod +w $out/nixos.img
+              # make-disk-image hardcodes the ESP FAT label to "ESP"; relabel the
+              # volume to ESP-USB so /boot (by-label/ESP-USB) can't resolve to an
+              # internal drive's ESP. Volume label only — bootloader files are
+              # untouched, and UEFI loads /EFI/BOOT/BOOTX64.EFI regardless.
+              espStart=$(parted -sm "$out/nixos.img" unit B print | awk -F: '$1==1 {gsub("B","",$2); print $2}')
+              echo "ESP partition starts at byte $espStart — relabelling to ESP-USB"
+              export MTOOLS_SKIP_CHECK=1
+              mlabel -i "$out/nixos.img@@$espStart" ::ESP-USB
+              printf 'verify ESP label: '; mlabel -i "$out/nixos.img@@$espStart" -s :: || true
+            '';
+
         # Backwards compat
         iso = self.nixosConfigurations.douro.config.system.build.isoImage;
       };

packages/hal/src/validators/ebds/ebds-rs232.ts

@@ -313,6 +313,13 @@ export class EbdsRs232 extends EventEmitter {
   private serial: SerialPort | null = null
   private ack: number = 0x0
   private enabledDenominations: number = 0x00
+  // Latched escrow decision. In EBDS the stack/return choice is NOT a one-shot
+  // message — it's carried as bits in the omnibus poll command, and the device
+  // holds the escrowed note until a poll asserts stack or return. We keep the
+  // action set and re-assert it on every poll until the device leaves escrow
+  // (cleared in _process), so a single dropped/collided frame no longer strands
+  // the note in escrow forever.
+  private pendingAction: 'none' | 'stack' | 'return' = 'none'
   private lastStatusFlags: string | null = null
   private firmwareLogged: boolean = false
 
@@ -405,23 +412,38 @@ export class EbdsRs232 extends EventEmitter {
 
   // -- Commands (Appendix D, Controller Message) ---------------------------
 
-  /** Send an Omnibus poll command with current denomination mask */
+  /**
+   * Command byte 1 for the omnibus poll, encoding any latched escrow action.
+   * `stack` (0x3f) and `return` (0x5f) differ from the plain poll (0x1b) only
+   * in the stack/return bits; while an action is latched every poll re-asserts
+   * it until the device acts.
+   */
+  private commandByte(): number {
+    if (this.pendingAction === 'stack') return 0x3f
+    if (this.pendingAction === 'return') return 0x5f
+    return 0x1b
+  }
+
+  /** Send an Omnibus poll command with the current mask + latched action */
   poll(): void {
-    this._dispatch([this.enabledDenominations, 0x1b, 0x10])
+    this._dispatch([this.enabledDenominations, this.commandByte(), 0x10])
   }
 
-  /** Stack the bill currently in escrow */
+  /** Latch "stack the escrowed note"; re-asserted each poll until it takes. */
   stack(): void {
-    this._dispatch([this.enabledDenominations, 0x3f, 0x10])
+    this.pendingAction = 'stack'
+    this.poll()
   }
 
-  /** Reject/return the bill currently in escrow */
+  /** Latch "return the escrowed note"; re-asserted each poll until it takes. */
   reject(): void {
-    this._dispatch([this.enabledDenominations, 0x5f, 0x10])
+    this.pendingAction = 'return'
+    this.poll()
   }
 
   /** Send initial setup command (disable all, reset state) */
   reset(): void {
+    this.pendingAction = 'none'
     this._dispatch([0x00, 0x1b, 0x10])
   }
 
@@ -470,7 +492,13 @@ export class EbdsRs232 extends EventEmitter {
       validatePacket(packet)
       const result = interpret(packet)
       if (result) {
-        if (result.destructedData) this._logStatusOnChange(result.destructedData)
+        if (result.destructedData) {
+          // Clear a latched stack/return once the device has left escrow — it
+          // is now stacking/returning/idle, so we must stop asserting the
+          // action or it would leak onto the next note.
+          if (!result.destructedData[0].escrowed) this.pendingAction = 'none'
+          this._logStatusOnChange(result.destructedData)
+        }
         this.emit('message', result)
       }
     } catch (ex) {