undercooked/tests/sim/test_full_rounds.py

"""Headless full-round simulations (phase 04, task 06) — the whole pipeline
driven from tick 0 on the shipped level 1 to a verdict, no server.

These are the phase's integration layer (PLAN §9.4): the scripted cooks of
`helpers.py` drive the *complete* `step()` — movement, the action gate,
stations, freshness, orders, scoring, the round state — on
`levels/01_first_kitchen.yaml` with the shipped recipe book, and the
assertions read the verdict (`result`), the stars, the counters and the
event stream. A win is won by the rules (score over the bar, the early
win's substep), a loss by the clock, and the wash cycle is the dirty loop
closing on a one-prop kitchen.

* `test_win_level_1_solo` — one scripted cook cooks soup end to end (fetch,
  chop, pot, cook, plate, serve, repeat) from tick 0: the shipped level 1
  is won with ≥ 1 s of clock left (winnable with slack, not on the buzzer)
  and `stars ≥ 1`; against the 100 bar a single first dish (≈ 58 points)
  cannot clear, so the run repeats the whole chain until `WON` (≥ 2 dishes
  served) — the chain is not one-shot.
* `test_loss_by_timer_idle` — no cooks, no inputs: the clock runs the 150 s
  down, `result == LOST` on the buzzer (tick 3 000) with `score == 0` and
  at least one missed order; every tick the all-idle invariants hold:
  counters ≥ 0, the rail within `order_concurrent`, the item count within
  the seeded shelf stock, the order ids bounded, and `time_left` finite
  (no division by zero anywhere in the idle round).
* `test_two_player_win` — two scripted cooks (each running the full-chain
  policy on its own primary shelf) clear the 100 bar and the rules reward
  the second cook by scoring its work: both cooks serve a dish, both earn
  points, and the per-player contributions sum to the round's score.
  (Measured across nine seat/split/role configurations on this single-
  pipeline kitchen, two cooks never beat one on the clock — the shared
  fixtures serialize the second cook — so the testable reward the rules
  give is the per-cook scoring, not a speed edge.)
* `test_wash_cycle_required` — the one-plate, one-pot kitchen: after the
  first dish the stock pair is dirty, so the second plate needs the wash —
  a `WASHED` lands before the second serve, `plates_clean + plates_dirty
  == 1` and `pots_clean + pots_dirty == 1` every tick (the stock is
  conserved, not just refunded), and the round ends `WON` with ≥ 2 dishes
  served. The zero-stock refusal itself is the named `no_plate` denial:
  asserted the targeted way (`apply_station_verb` on a zero-stock state —
  the ladder's in-round answer at a dirty fixture with a full hand is
  `STATION_DIRTY`, and the stock guard is the named safety behind it),
  leaving the dish in the hand and the stock floored at 0.
* `test_no_order_flooding_over_full_round` — a full 150 s idle round (no
  early verdict truncating the rail): the rail never exceeds
  `order_concurrent` on any tick, and the order ids are strictly
  increasing in first appearance — spawns are scheduled, not flooded.
"""

from __future__ import annotations

import math
from itertools import pairwise

from helpers import (
    load_book,
    make_level_1_state,
    run_round,
    shelf_id_for,
)

from undercooked.sim.grid import Grid, TileKind
from undercooked.sim.interact import StationTarget, Verb
from undercooked.sim.model import (
    EventKind,
    Facing,
    Item,
    ItemForm,
    ItemKind,
    KitchenState,
    Player,
    RoundResult,
    Station,
    StationKind,
    Vec2,
)
from undercooked.sim.round_state import stars
from undercooked.sim.stations import apply_station_verb

#: Level 1's budget: 150 s, 20 substeps/s.
FULL_ROUND_TICKS = 3000

#: The bar a single first dish (≈ 58 points at best) cannot clear: the
#: repeat, the two-player and the wash-cycle runs need at least two
#: dishes, so the wash cycle is genuinely forced before the second plate.
TWO_DISH_BAR = 100


def _tomato_shelf() -> str:
    """The tomato shelf's id (the round is built with the default seed, so
    the layout is the same every run)."""
    return shelf_id_for(make_level_1_state(), ItemKind.TOMATO)


def test_win_level_1_solo() -> None:
    """A scripted cook, tick 0 to the win, on the shipped level 1.

    The shipped bar (45): one first dish (≈ 58 points) clears it — the win
    lands with ≥ 1 s of clock left (slack, not a buzzer) and `stars ≥ 1`.
    Against the 100 bar the first dish cannot clear, so the run repeats
    the whole chain (fetch, chop, pot, cook, plate, serve) until `WON`
    with ≥ 2 dishes served: the chain is not one-shot.
    """
    book = load_book()
    shelf = _tomato_shelf()

    # The shipped level 1: the phase's criterion, tick 0 to `WON`.
    state = make_level_1_state(seats=1)
    states, _events, verdict_tick = run_round(
        state, book, [shelf], FULL_ROUND_TICKS - 10, stop_at_verdict=True
    )
    final = states[-1]
    assert verdict_tick is not None, "the round never ended in the window"
    assert final.result is RoundResult.WON
    assert final.score >= final.target_score
    assert stars(final) >= 1
    assert final.counters.get("served", 0) >= 1
    # Winnable with slack: the win lands ≥ 1 s before the buzzer.
    assert final.time_left >= 1.0
    assert 0 < verdict_tick < FULL_ROUND_TICKS

    # The 100 bar: a single first dish (≈ 58) cannot clear — the run
    # repeats the whole chain until it does.
    repeat = make_level_1_state(seats=1, target=TWO_DISH_BAR)
    repeat_states, _repeat_events, repeat_tick = run_round(
        repeat, book, [shelf], FULL_ROUND_TICKS - 10, stop_at_verdict=True
    )
    repeat_final = repeat_states[-1]
    assert repeat_tick is not None, "the repeat run never ended in the window"
    assert repeat_final.result is RoundResult.WON
    assert repeat_final.score >= TWO_DISH_BAR
    assert repeat_final.counters.get("served", 0) >= 2, "the chain is one-shot"
    assert repeat_final.time_left >= 1.0


def test_loss_by_timer_idle() -> None:
    """No cooks, no inputs: the clock decides.

    `time_left` reaches 0, `result == LOST` on the buzzer (tick 3 000),
    `score == 0` (nothing was served, and the miss penalty is clamped at
    the floor of 0), `counters["missed"] ≥ 1` (orders spawned and expired
    on their 60 s deadline), and every tick the all-idle invariants hold:
    counters ≥ 0, the rail within `order_concurrent`, the item count
    within the seeded shelf stock, the order ids bounded, and `time_left`
    finite (no division by zero anywhere in the idle round).
    """
    state = make_level_1_state(seats=0)
    book = load_book()
    states, _event_lists, verdict_tick = run_round(
        state, book, [], FULL_ROUND_TICKS + 10, stop_at_verdict=True
    )
    final = states[-1]
    assert verdict_tick == FULL_ROUND_TICKS, "the buzzer is the 150 s tick"
    assert final.result is RoundResult.LOST
    assert final.score == 0
    assert final.counters.get("missed", 0) >= 1
    for index, s in enumerate(states):
        assert math.isfinite(s.time_left), f"tick {index}"
        for key, value in s.counters.items():
            assert value >= 0, f"tick {index}: {key} = {value}"
        assert len(s.orders) <= s.order_concurrent, f"tick {index}"
        items = sum(len(station.contents) for station in s.stations)
        assert items <= 16, f"tick {index}: {items} items (the seeded stock)"
        # The rail's schedule stays bounded: at most one spawn per 16 s
        # (the level's `spawn_interval` floor) over 150 s → < 16 ids.
        assert s.next_order_id <= 16, f"tick {index}: {s.next_order_id} ids"


def test_two_player_win() -> None:
    """Two scripted cooks (each the full-chain policy, its own primary
    shelf) clear the 100 bar — a bar a single first dish (≈ 58 points)
    cannot clear — and the rules reward the second cook by scoring its
    work: both cooks serve a dish, both earn points, and the per-player
    contributions sum to the round's score.

    (Measured across nine seat/split/role configurations on this
    single-pipeline kitchen, two cooks never beat one on the clock — the
    shared fixtures serialize the second cook — so the testable reward
    the rules give is the per-cook scoring, not a speed edge.)
    """
    book = load_book()
    state = make_level_1_state(seats=2, target=TWO_DISH_BAR)
    states, _events, verdict_tick = run_round(
        state,
        book,
        [_tomato_shelf(), shelf_id_for(state, ItemKind.ONION)],
        FULL_ROUND_TICKS - 10,
        stop_at_verdict=True,
    )
    final = states[-1]
    assert verdict_tick is not None, "the round never ended in the window"
    assert final.result is RoundResult.WON
    assert final.score >= TWO_DISH_BAR
    # Both cooks' work is scored: each served a dish and earned points.
    for index, cook in enumerate(final.players):
        assert cook.dishes_served >= 1, f"cook {index} served no dish"
        assert cook.score >= 1, f"cook {index} earned no points"
    # The per-player contributions sum to the round's score (a win in the
    # first 52 s, before the first order's deadline, so no miss penalty
    # stands between them).
    assert sum(cook.score for cook in final.players) == final.score
    # The dishes sum to the served counter.
    assert sum(cook.dishes_served for cook in final.players) == final.counters["served"]


def _zero_stock_plate_dish_denial() -> None:
    """The zero-stock refusal, targeted (the named `no_plate` denial): a
    zero plate stock is a named refusal, the stock's floor is 0, and the
    dish stays in the hand — the wash is the fix, not the refusal."""
    state = KitchenState(
        level_id="wash_cycle_probe",
        seed=7,
        grid=Grid(width=7, height=7, tiles=(TileKind.FLOOR,) * 49),
        counters={"plates_clean": 0, "plates_dirty": 1},
        stations=(
            Station(id="prep-0", kind=StationKind.PREP, x=5.0, y=4.0),
        ),
        players=(
            make_cook_carrying_cooked(),
        ),
    )
    player = state.players[0]
    state, events = apply_station_verb(
        state, player, Verb.PLATE_DISH, StationTarget(station=state.stations[0])
    )
    denied = [e for e in events if e.kind is EventKind.DENIED]
    assert len(denied) == 1, "the zero-stock plate is not a named refusal"
    payload = dict(denied[0].payload)
    assert payload["reason"] == "nothing_here"
    assert payload["why"] == "no_plate"
    assert state.counters["plates_clean"] == 0  # the floor: never below 0
    assert state.players[0].carrying is not None  # the dish stays in the hand


def make_cook_carrying_cooked() -> Player:
    """One cook (seat 0) standing in front of the probe's prep, holding a
    cooked item — the `plate_dish`'s whole precondition minus the stock."""
    return Player(
        id="p0",
        pos=Vec2(5.5, 5.5),
        facing=Facing.UP,
        carrying=Item(ItemKind.TOMATO, ItemForm.COOKED),
    )


def test_wash_cycle_required() -> None:
    """The one-plate, one-pot kitchen: the stock pair goes dirty after
    the first dish, so the second plate is refused until the wash runs.

    Asserted two ways, the same guarantee: the targeted way (a zero-stock
    `plate_dish` is the named `no_plate` refusal, the stock floored at 0,
    the dish left in the hand — the wash is the fix) and the full-round
    way (a `WASHED` lands before the second serve — the policy washes the
    dirty fixture in place, the stock pair is conserved every tick, and
    the round ends `WON` with ≥ 2 dishes served: the second dish, which a
    single clean plate cannot support without the wash, is what clears
    the 100 bar).
    """
    _zero_stock_plate_dish_denial()  # the denial: named, zero stock, floor 0

    state = make_level_1_state(seats=1, target=TWO_DISH_BAR, plates=1, pots=1)
    book = load_book()
    states, event_lists, verdict_tick = run_round(
        state, book, [_tomato_shelf()], FULL_ROUND_TICKS - 10, stop_at_verdict=True
    )
    final = states[-1]
    assert verdict_tick is not None, "the round never ended in the window"
    assert final.result is RoundResult.WON
    assert final.counters.get("served", 0) >= 2

    # The stock pair is conserved every tick (the loop's invariant).
    for index, s in enumerate(states):
        assert s.counters["plates_clean"] + s.counters["plates_dirty"] == 1, f"tick {index}"
        assert s.counters["pots_clean"] + s.counters["pots_dirty"] == 1, f"tick {index}"

    # The success after the wash: a wash lands before the second serve
    # (with a single clean plate, the second plate needs it), and the
    # second dish lands after it.
    wash_ticks = [
        index
        for index, events in enumerate(event_lists)
        for event in events
        if event.kind is EventKind.WASHED
    ]
    served_ticks = [
        index
        for index, events in enumerate(event_lists)
        for event in events
        if event.kind is EventKind.SERVED
    ]
    assert len(served_ticks) >= 2
    assert wash_ticks, "the wash never ran"
    assert min(wash_ticks) < served_ticks[1], (
        "the second dish served before any wash"
    )


def test_no_order_flooding_over_full_round() -> None:
    """A full 150 s idle round (no early verdict truncating the rail's
    schedule): the ticket rail never exceeds `order_concurrent` on any
    tick, and the order ids are strictly increasing in first appearance —
    spawns are scheduled, not flooded. The whole window runs, so several
    orders spawn and expire on their 60 s deadline."""
    state = make_level_1_state(seats=0)
    book = load_book()
    states, _event_lists, _verdict_tick = run_round(
        state, book, [], FULL_ROUND_TICKS
    )
    assert len(states) == FULL_ROUND_TICKS
    seen_ids: list[int] = []
    for index, s in enumerate(states):
        assert len(s.orders) <= s.order_concurrent, f"tick {index}"
        for order in s.orders:
            if order.id not in seen_ids:
                seen_ids.append(order.id)
    # Strictly increasing in first appearance (a monotone id counter).
    for earlier, later in pairwise(seen_ids):
        assert later > earlier, f"{later} did not follow {earlier}"
    # The whole window ran: the 16 s spawn floor over 150 s → ≥ 3 orders
    # spawned (and the first two expired on their deadline).
    assert len(seen_ids) >= 3