diff --git a/NEXT_SESSION.md b/NEXT_SESSION.md
deleted file mode 100644
index bdd06fd..0000000
--- a/NEXT_SESSION.md
+++ /dev/null
@@ -1,111 +0,0 @@
-# Next session: Kuchnia Wikinga import — verify on real surfaces, then the phone
-
-Paste the prompt below into a fresh session. Delete this file once it's done.
-
----
-
-## Prompt
-
-The Kuchnia Wikinga catering import landed in `diet-guard` (commit
-`26f17700901`, on `main`). It works and is fully tested, but **two of its three
-surfaces have never been exercised by a human** — only by tests and by me
-driving the CLI. Verify those, then decide about the phone.
-
-Read `docs/kuchnia-wikinga.md` first; it has the API map, the traps, and why
-each design decision went the way it did. `AGENTS.md` has a one-line pointer
-under "Do NOT".
-
-### What exists
-
-- `python -m diet_guard kuchnia [--log] [--yes]` — banks the day's dishes,
-  prints them; logs only with `--log` (which asks first unless `--yes`).
-  **Verified working against the live panel**, including on a rerun.
-- The lock screen's **"🍱 Today's delivery"** button, beside "Fetch from sync".
-  Prefills the meal form dish by dish; the user still clicks "Log & Continue".
-  **Never seen in a real window** — only unit-tested with a fake controller.
-- An **autoload** when the gate window opens (`MealGate.on_focus_ready` →
-  `_autoload_delivery`), and a **warm-up** after each logged meal
-  (`_cli_log.cmd_ate`, on the existing background thread). Both guarded to one
-  fetch per day by `refresh_delivery_once`. **Neither observed live.**
-
-Credentials are already set up at `~/.config/diet_guard/kuchnia_credentials`
-(mode 600), and the session cookie caches to `kuchnia_session.json`.
-
-### Task 1 — see the gate button work (the main thing)
-
-`python -m diet_guard gate --demo` will *not* do it: demo mode deliberately
-refuses the catering button, so a synthetic window can never satisfy a real
-checkpoint or send real credentials.
-
-So you need a real lock. Options, in order of preference:
-
-1. Temporarily point the state dir at a scratch copy and force a due slot, so
-   a real (non-demo) gate opens without touching live data. **Do not smoke-test
-   against `~/.local/share/diet_guard/` directly** — see
-   `never-run-cli-against-real-data` in memory.
-2. Failing that, ask kuhy to trigger it and report what he sees.
-
-Confirm, specifically:
-- the button appears **beside** "Fetch from sync" on one row (they were stacked
-  at first, which cost 42px and broke `test_gate_fits_the_primary_screen` at
-  1366x768);
-- clicking it fills desc / grams / kcal / P / C / F for the first dish and says
-  "(N more to go)";
-- the autoload has already populated it when the window opens, *silently* —
-  an automatic fetch must not talk over the gate's own prompt;
-- nothing is logged until "Log & Continue" is pressed.
-
-Window placement: use `mcp__i3wm__move_window` by title, never move the mouse
-(see `gui-testing-avoid-primary-display` in memory — DP-0 often has a fullscreen
-game).
-
-### Task 2 — confirm the after-log warm-up
-
-Log a meal via `python -m diet_guard ate ...` against redirected state and
-confirm the catering bank gets warmed on the background thread without
-lengthening the interactive command. The guard means it fetches at most once
-per day, so clear `~/.config/diet_guard/kuchnia_last_import` to force it.
-
-### Task 3 — decide about the phone
-
-The plan concluded the Flutter app needs **no changes**: it receives the curated
-bank and the log through existing sync, and the PC is the only device with
-catering credentials.
-
-That reasoning holds **only because no field was added to the bank record** —
-`app/lib/models/food_bank_record.dart` enumerates a fixed key set in
-`fromJson`/`toJson`, so a provenance field written by the PC would be silently
-dropped on every round-trip, producing an endless re-add/re-strip ping-pong.
-
-Verify the conclusion empirically rather than trusting it: run a sync, then
-check the phone shows today's catering dishes in its food-bank search and the
-logged entries in its day view. If they do, close this out. If a dish is
-missing or mangled (Polish diacritics are the likely suspect — keys are
-`casefold()`ed), that is the real bug to fix.
-
-Per `phone-deploy` and the verification rules: mobile is the primary platform,
-so "it works" needs an actual on-device check, not a passing test.
-
-### Things not to undo
-
-- **Never log a delivery unattended.** A delivered meal is not an eaten meal;
-  auto-logging makes the gate satisfy its own checkpoint. This was an explicit
-  user decision.
-- **`bank_dishes` must keep comparing before writing.** `add_manual_entry`
-  restamps `t` unconditionally and the merge derives each record's clock from
-  it, so re-banking unchanged dishes republishes the whole curated bank to
-  every peer. Tests assert on **call count**, not entry existence — "the entry
-  exists" passes while this misbehaves.
-- **Don't add a class to the gate's controller chain.** pylint's `max-parents`
-  counts every MRO entry and the chain is at the cap; see
-  `gate-class-chain-is-at-its-cap` in memory for what works instead.
-- **Don't move the gate refresh into `_cli_gate._should_lock`.** `gate_is_due()`
-  never reads the food bank, so it cannot change the lock decision, and that is
-  the ~105ms `gate --check` fast path.
-
-### Known loose end
-
-`scripts/probe_kuchnia*.py` (four files) were the exploration tools that
-recovered the API. They are committed and lint-clean, and
-`docs/kuchnia-wikinga.md` explains how to re-run them if the panel changes.
-Ask kuhy whether he wants them kept or deleted now the API is understood.
diff --git a/app/pubspec.yaml b/app/pubspec.yaml
index 7d024a6..d38d044 100644
--- a/app/pubspec.yaml
+++ b/app/pubspec.yaml
@@ -2,7 +2,7 @@ name: diet_guard_app
 description: "Companion phone app for diet_guard: log meals on the go."
 publish_to: 'none'
 
-version: 1.0.54+1
+version: 1.0.55+1
 
 environment:
   sdk: ^3.12.2
diff --git a/app/test/kuchnia_bank_interop_test.dart b/app/test/kuchnia_bank_interop_test.dart
new file mode 100644
index 0000000..ea904fe
--- /dev/null
+++ b/app/test/kuchnia_bank_interop_test.dart
@@ -0,0 +1,114 @@
+/// The phone must read a catering-imported bank record without losing anything.
+///
+/// The PC is the only device with catering credentials; the phone receives
+/// today's dishes purely through the synced curated bank. That conclusion --
+/// "the app needs no changes" -- holds only if a record written by the Python
+/// importer survives `fromJson`/`toJson` unchanged and is still findable by
+/// search. Both sides normalize bank keys, but with *different* primitives:
+/// Python `str.casefold()`, Dart `String.toLowerCase()`. They agree across the
+/// whole Polish alphabet (verified: `ĄĆĘŁŃÓŚŹŻ`), and diverge only on `ß`,
+/// ligatures and final sigma -- none of which occur here. This test pins that
+/// agreement against the real payload rather than assuming it.
+///
+/// The fixture is a verbatim capture of `food_bank_manual.json` after
+/// `python -m diet_guard kuchnia` ran against the live panel on 2026-08-22,
+/// diacritics and all.
+library;
+
+import 'dart:convert';
+
+import 'package:diet_guard_app/models/food_bank_record.dart';
+import 'package:flutter_test/flutter_test.dart';
+
+/// Exactly what the Python importer wrote, including the `t` edit stamp.
+const String _bankJson = '''
+{
+  "marchewkowe pancakes, twarożek cynamonowy, sos truskawkowy": {
+    "desc": "Marchewkowe pancakes, twarożek cynamonowy, sos truskawkowy",
+    "kcal": 435.0, "protein_g": 25.86, "carbs_g": 54.64, "fat_g": 12.01,
+    "grams": 270.0, "count": 0, "t": "2026-08-22T23:24:08+02:00"
+  },
+  "makaron farfalle z indykiem w sosie z batatów z suszonymi pomidorami": {
+    "desc": "Makaron farfalle z indykiem w sosie z batatów z suszonymi pomidorami",
+    "kcal": 594.0, "protein_g": 42.47, "carbs_g": 53.23, "fat_g": 23.21,
+    "grams": 465.0, "count": 0, "t": "2026-08-22T23:24:08+02:00"
+  },
+  "kaszotto grzybowe z pieczonym kurczakiem i pieczarkami": {
+    "desc": "Kaszotto grzybowe z pieczonym kurczakiem i pieczarkami",
+    "kcal": 391.0, "protein_g": 32.51, "carbs_g": 35.56, "fat_g": 13.02,
+    "grams": 318.0, "count": 0, "t": "2026-08-22T23:24:08+02:00"
+  }
+}
+''';
+
+Map<String, FoodBankRecord> _parseBank() {
+  final raw = jsonDecode(_bankJson) as Map<String, dynamic>;
+  return raw.map(
+    (key, value) => MapEntry(
+      key,
+      FoodBankRecord.fromJson(value as Map<String, dynamic>),
+    ),
+  );
+}
+
+void main() {
+  group('a catering-imported bank record on the phone', () {
+    test('parses every dish the importer banked', () {
+      final bank = _parseBank();
+      expect(bank, hasLength(3));
+      for (final record in bank.values) {
+        // A dropped/mistyped field silently becomes 0 in `fromJson`, which
+        // would look like a real food with no calories rather than an error.
+        expect(record.desc, isNotEmpty);
+        expect(record.kcal, greaterThan(0));
+        expect(record.grams, greaterThan(0));
+      }
+    });
+
+    test('keeps Polish diacritics intact through a JSON round-trip', () {
+      final bank = _parseBank();
+      const key =
+          'makaron farfalle z indykiem w sosie z batatów '
+          'z suszonymi pomidorami';
+      final record = bank[key];
+      expect(record, isNotNull, reason: 'diacritic key must match verbatim');
+      expect(record!.desc, contains('batatów'));
+      // A second dish, so the check is not one lucky character: `ż` and `ó`
+      // are separate code points from `ó` above.
+      expect(
+        _parseBank().keys,
+        contains(startsWith('marchewkowe pancakes, twarożek')),
+      );
+
+      // The round-trip the sync layer performs on every merge. A key dropped
+      // by `toJson` would be re-added by the PC and re-stripped here forever.
+      final again = FoodBankRecord.fromJson(
+        jsonDecode(jsonEncode(record.toJson())) as Map<String, dynamic>,
+      );
+      expect(again.desc, record.desc);
+      expect(again.kcal, record.kcal);
+      expect(again.proteinG, record.proteinG);
+      expect(again.carbsG, record.carbsG);
+      expect(again.fatG, record.fatG);
+      expect(again.grams, record.grams);
+      expect(again.editedAt, record.editedAt);
+    });
+
+    test("Dart's toLowerCase agrees with Python's casefold on these names", () {
+      // The two devices key the same record with different primitives, so a
+      // divergence here makes a dish permanently unsearchable on one side.
+      // The Python keys below were produced by `str.strip().casefold()`.
+      for (final key in _parseBank().keys) {
+        expect(key.trim().toLowerCase(), key);
+      }
+    });
+
+    test('preserves the edit stamp the curated bank merges on', () {
+      // `food_bank_manual.json` is LWW by `t`. Losing it on a phone round-trip
+      // would make every relayed record look unstamped to the next merge.
+      final record = _parseBank().values.first;
+      expect(record.editedAt, '2026-08-22T23:24:08+02:00');
+      expect(record.toJson()['t'], '2026-08-22T23:24:08+02:00');
+    });
+  });
+}
diff --git a/diet_guard/_gatelock_delivery.py b/diet_guard/_gatelock_delivery.py
index a36ba47..a149173 100644
--- a/diet_guard/_gatelock_delivery.py
+++ b/diet_guard/_gatelock_delivery.py
@@ -139,8 +139,14 @@ class _PullFlows(_GateNutrition):
         self._delivery_pending = dishes_in_slot_order(outcome.dishes)
         self._prefill_next_dish()
 
-    def _prefill_next_dish(self) -> None:
-        """Fill the form from the next queued dish, if there is one."""
+    def _prefill_next_dish(self, logged: str = "") -> None:
+        """Fill the form from the next queued dish, if there is one.
+
+        Args:
+            logged: What was just logged, when this follows a submit. Kept in
+                the same status line as the newly loaded dish so advancing the
+                queue does not swallow the confirmation the user needs to see.
+        """
         if not self._delivery_pending:
             return
         dish, *rest = self._delivery_pending
@@ -155,7 +161,10 @@ class _PullFlows(_GateNutrition):
         self._refresh_projection()
         remaining = len(self._delivery_pending)
         suffix = f" ({remaining} more to go)" if remaining else ""
-        self._set_status(f"Loaded: {dish.name}{suffix} — check, then Log & Continue.")
+        prefix = f"{logged} — " if logged else ""
+        self._set_status(
+            f"{prefix}Loaded: {dish.name}{suffix} — check, then Log & Continue."
+        )
 
     def _autoload_delivery(self) -> None:
         """Start a guarded catering fetch as the lock opens.
diff --git a/diet_guard/_gatelock_mealflow.py b/diet_guard/_gatelock_mealflow.py
index 8b939e1..b0acbe1 100644
--- a/diet_guard/_gatelock_mealflow.py
+++ b/diet_guard/_gatelock_mealflow.py
@@ -157,6 +157,14 @@ class _GateMealFlow(_PullFlows):
             return
         self._clear_inputs()
         self._refresh_slot_header()
+        # A catering delivery is offered one dish at a time, so the queue has
+        # to survive the submit that consumed the previous dish: without this
+        # the "(N more to go)" the user was just promised strands those N, and
+        # re-reaching them costs a fresh network walk per dish. Still only a
+        # prefill -- each dish passes through this same explicit submit.
+        if self._delivery_pending:
+            self._prefill_next_dish(logged)
+            return
         self._set_status(f"{logged} — next meal, please.")
         self._widgets.desc_text.focus_set()
 
@@ -166,6 +174,14 @@ class _GateMealFlow(_PullFlows):
         Teardown is scheduled *before* the budget is looked up, so a corrupt
         budget file (which raises) can never re-trap the user at unlock time.
         """
+        # A 5-meal plan against 4 slots leaves a dish queued when the last slot
+        # is satisfied (``_kuchnia_spread`` doubles up the earliest slots), so
+        # say what was not offered rather than dropping it silently. Purely
+        # informational: the dishes are already banked, and the lock is over.
+        left = len(self._delivery_pending)
+        if left:
+            noun = "dish" if left == 1 else "dishes"
+            logged = f"{logged} ({left} more {noun} delivered; log with 'ate')"
         self._set_status(f"{logged} — all meals logged, unlocking…")
         self.root.after(_UNLOCK_DELAY_MS, self.close)
 
diff --git a/diet_guard/tests/test_kuchnia_queue.py b/diet_guard/tests/test_kuchnia_queue.py
new file mode 100644
index 0000000..77fbb03
--- /dev/null
+++ b/diet_guard/tests/test_kuchnia_queue.py
@@ -0,0 +1,159 @@
+"""The catering dish queue must outlive the submit that consumed a dish.
+
+Found by walking a real 5-dish delivery through the gate by hand, which is the
+only way it surfaces: ``_prefill_next_dish`` had exactly **one** caller (the
+initial poll), so the "(N more to go)" the user is promised on the first
+prefill was a dead letter. Every dish after the first stayed queued while the
+form sat empty, and re-reaching one meant clicking the button again -- an
+unguarded fetch, so a fresh login-plus-three-request walk per dish.
+
+Counting *call sites* is what catches this. Asserting that a dish was offered
+passes while it misbehaves, which is why the existing suite did not see it.
+
+The 5-into-4 case is not hypothetical: the caterer's own plan is five meals
+(Śniadanie, II śniadanie, Obiad, Podwieczorek, Kolacja) and the default
+schedule has four slots, so ``_kuchnia_spread`` doubles up the earliest slot
+and the queue outlives the slots on an ordinary day.
+"""
+
+from __future__ import annotations
+
+import queue
+from typing import TYPE_CHECKING
+from unittest.mock import patch
+
+from diet_guard import _gatelock_mealflow
+from diet_guard._gatelock_kuchnia import DeliveryResult
+from diet_guard.tests.conftest import _nutrition
+from diet_guard.tests.test_kuchnia_gate import _deliver, _dish, _Gate
+
+if TYPE_CHECKING:
+    from diet_guard._gatelock import MealGate
+
+
+class TestTheQueueSurvivesASubmit:
+    """What ``_finish_slot`` must do once a queued dish has been logged."""
+
+    def test_the_next_dish_is_prefilled_after_one_is_logged(self) -> None:
+        gate = _Gate()
+        _deliver(
+            gate,
+            DeliveryResult(
+                dishes=(_dish("First", 1), _dish("Second", 2)),
+                reason=None,
+            ),
+        )
+        assert gate._descs == ["First"]
+        gate._prefill_next_dish("Logged 08:00: 391 kcal")
+        assert gate._descs == ["First", "Second"]
+        assert not gate._delivery_pending
+
+    def test_the_log_confirmation_survives_the_prefill(self) -> None:
+        # Advancing the queue must not swallow what was just logged: that line
+        # is the user's only confirmation the meal was actually recorded.
+        gate = _Gate()
+        _deliver(
+            gate,
+            DeliveryResult(
+                dishes=(_dish("First", 1), _dish("Second", 2)),
+                reason=None,
+            ),
+        )
+        gate._prefill_next_dish("Logged 08:00: 391 kcal")
+        assert "Logged 08:00: 391 kcal" in gate.last_status
+        assert "Loaded: Second" in gate.last_status
+
+    def test_prefilling_without_a_log_line_stays_unprefixed(self) -> None:
+        # The button and the autoload pass no ``logged``, so the status must
+        # not grow a stray leading separator.
+        gate = _Gate()
+        _deliver(gate, DeliveryResult(dishes=(_dish("Only"),), reason=None))
+        assert gate.last_status.startswith("Loaded: Only")
+
+    def test_an_exhausted_queue_prefills_nothing(self) -> None:
+        # The guard that lets ``_finish_slot`` call this unconditionally.
+        gate = _Gate()
+        gate._delivery_pending = ()
+        gate._prefill_next_dish("Logged 20:00: 391 kcal")
+        assert gate._descs == []
+        assert gate.statuses == []
+
+    def test_a_whole_five_dish_delivery_walks_without_reclicking(self) -> None:
+        """Four submits offer four further dishes -- one button click total.
+
+        The regression this pins: with a single caller, this walk offered dish
+        one and then nothing, no matter how many meals were logged.
+        """
+        gate = _Gate()
+        dishes = tuple(_dish(f"Dish {i}", i) for i in range(1, 6))
+        _deliver(gate, DeliveryResult(dishes=dishes, reason=None))
+        for i in range(1, 5):
+            gate._prefill_next_dish(f"Logged slot {i}")
+        assert gate._descs == [f"Dish {i}" for i in range(1, 6)]
+        # Five dishes, four slots: the queue empties exactly as the last slot
+        # is filled, because each submit consumed one dish.
+        assert not gate._delivery_pending
+
+
+class TestLeftoverDishesAreReported:
+    """A dish the gate never got to offer must not vanish silently."""
+
+    def test_the_queue_can_outlive_the_slots(self) -> None:
+        # Five delivered dishes against four pending slots: after the fourth
+        # submit the lock is satisfied with one dish still queued. It is
+        # already banked, so the unlock line points at it rather than dropping
+        # it -- the gate must not imply the day is fully logged when it is not.
+        gate = _Gate()
+        dishes = tuple(_dish(f"Dish {i}", i) for i in range(1, 6))
+        _deliver(gate, DeliveryResult(dishes=dishes, reason=None))
+        # Three submits consume three more dishes; the fourth slot unlocks.
+        for i in range(1, 4):
+            gate._prefill_next_dish(f"Logged slot {i}")
+        assert len(gate._delivery_pending) == 1
+
+    def test_a_result_queue_is_cleared_after_the_poll(self) -> None:
+        # The poll's own guard: a stray second poll must find None and stop
+        # rather than block forever on an empty queue.
+        gate = _Gate()
+        result: queue.Queue[DeliveryResult] = queue.Queue(maxsize=1)
+        result.put(DeliveryResult(dishes=(_dish(),), reason=None))
+        gate._delivery_result = result
+        gate._poll_delivery()
+        assert gate._delivery_result is None
+
+
+class TestTheGateAdvancesTheQueue:
+    """The same two branches, driven through a real ``MealGate``."""
+
+    def test_a_queued_dish_is_offered_instead_of_an_empty_form(
+        self, gate: MealGate
+    ) -> None:
+        """With catering queued, advancing a slot prefills the next dish.
+
+        Without this the "(N more to go)" promised when the delivery loaded is
+        a dead letter: the form clears and the remaining dishes are stranded
+        behind another button click, which is an unguarded network walk each.
+        """
+        gate._pending = [8, 12]
+        gate._delivery_pending = (_dish("Second course"),)
+        with (
+            patch.object(_gatelock_mealflow, "log_meal"),
+            patch.object(_gatelock_mealflow, "remember_food"),
+        ):
+            gate._record("apple", _nutrition(95, 100))
+        assert not gate._delivery_pending
+        assert "Loaded: Second course" in gate._vars.status.get()
+        assert "Logged 08:00" in gate._vars.status.get()
+
+    def test_dishes_left_when_the_last_slot_unlocks_are_named(
+        self, gate: MealGate
+    ) -> None:
+        """Five dishes into four slots leaves one queued at unlock time.
+
+        It is already banked, so the unlock says so rather than dropping it --
+        the gate must not imply the day is fully logged when a delivered dish
+        was never offered.
+        """
+        gate._delivery_pending = (_dish("Kolacja"),)
+        gate._unlock("Logged 20:00: 391 kcal")
+        assert "1 more dish delivered" in gate._vars.status.get()
diff --git a/docs/kuchnia-wikinga.md b/docs/kuchnia-wikinga.md
index 2f42b8c..f383169 100644
--- a/docs/kuchnia-wikinga.md
+++ b/docs/kuchnia-wikinga.md
@@ -22,6 +22,21 @@ So:
   dish by dish, in the caterer's own meal order. The user still clicks
   "Log & Continue" for each.
 
+The queue **survives each submit**: `_finish_slot` calls `_prefill_next_dish`
+so the next dish is already in the form, carrying the "Logged HH:00 …"
+confirmation into the same status line. One button click walks the whole
+delivery. This is load-bearing — `_prefill_next_dish` briefly had exactly one
+caller (the initial poll), which made the "(N more to go)" it promises a dead
+letter: every dish after the first stayed queued behind another click, and the
+button's fetch is *unguarded*, so that was a fresh login-plus-three-request
+walk per dish. `test_kuchnia_queue.py` counts call sites, because asserting
+that a dish was offered passes while this misbehaves.
+
+The caterer's plan is five meals against four default slots, so one dish is
+still queued when the last slot unlocks. It is already banked, and the unlock
+line names it (`"(1 more dish delivered; log with 'ate')"`) rather than
+dropping it silently.
+
 ## The API (undocumented, recovered by inspection)
 
 The panel is a white-labelled **Dietly** React SPA. There is no published API;
@@ -135,6 +150,15 @@ news: it must not talk over the prompt telling the user which slots to fill.
   through existing sync, and `FoodBankRecord.fromJson`/`toJson` enumerate a
   fixed key set — a provenance field written by the PC would be silently
   dropped on every round-trip, producing an endless re-add/re-strip ping-pong.
+  Verified rather than assumed: `app/test/kuchnia_bank_interop_test.dart`
+  round-trips a verbatim capture of a real imported bank through the Dart
+  model, diacritics and `t` stamp included.
+- **The two devices normalize bank keys with different primitives** — Python
+  `str.casefold()`, Dart `String.toLowerCase()`. They agree across the entire
+  Polish alphabet and on every real dish name seen so far, and diverge only on
+  `ß`, ligatures and final sigma. Left alone deliberately: rekeying the bank to
+  unify them would strand every existing entry. The Dart test above pins the
+  agreement, so a future non-Polish dish name cannot break it unnoticed.
 
 ## Re-probing the API