o2o-negosium-original/agent/tests/test_p2_state_reward_mapper.py

"""P2 검증 (계획서 P2 _검증_, 클린룸 반영).

기준 변경: "신규 결과 == Chat_server 결과"(독점 복제)가 아니라
"동일 config + 동일 입력 → 결정론적 동일 출력 + config 주입이 실제로 반영"으로 검증한다.

1. build_state/state_index 결정론 + 범위 [0, state_space_size).
2. mixed-radix 인코딩이 전 차원조합에 대해 [0,162) 전단사(bijection).
3. config 주입 효과: 테넌트별 임계값이 다르면 같은 입력이 다른 상태로 분류된다.
4. reward 결정론 + config(failure_penalty 등) 반영.
5. ActionCardMapper 라운드트립 + 중복방지 마스킹.
"""

import itertools
import os

import numpy as np

from negotiation.cards.action_card_mapper import ActionCardMapper
from negotiation.qtable.domain.model.snapshot import NegotiationOutcome, NegotiationSnapshot
from negotiation.qtable.domain.model.state import encode_index
from negotiation.qtable.domain.service.reward_calculator import RewardCalculator
from negotiation.qtable.domain.service.state_calculator import build_state, state_dims, state_index
from tenancy.config_loader import TenantConfigLoader

_TENANTS_DIR = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "tenants")


def _cfg(tid: str):
    return TenantConfigLoader(tenants_dir=_TENANTS_DIR, cache_ttl_seconds=0).load(tid)


def _snapshot(**over) -> NegotiationSnapshot:
    base = dict(
        revenue_amount=5_000_000,
        distribution_code="A",
        partner_count=1,
        acceptance_ratio=0.05,
        input_price=9800,
        anchor_price=9900,   # KT 앵커링가 (anchor < target)
        target_price=10000,  # KT 목표 매입가
        round_number=1,
        outcome=NegotiationOutcome.ONGOING,
    )
    base.update(over)
    return NegotiationSnapshot(**base)


def test_build_state_deterministic_and_in_range():
    cfg = _cfg("_base")
    snap = _snapshot()
    s1 = build_state(snap, cfg.state)
    s2 = build_state(snap, cfg.state)
    assert s1 == s2  # 결정론
    idx = state_index(snap, cfg.state)
    assert idx == state_index(snap, cfg.state)
    assert 0 <= idx < cfg.state.state_space_size


def test_encode_index_known_example():
    # dims=[3,3,3,3,2], 인덱스 (1,0,0,0,1) → ((((1)*3+0)*3+0)*3+0)*2+1 = 54*1 +1 = 55
    assert encode_index([1, 0, 0, 0, 1], [3, 3, 3, 3, 2]) == 55
    # 최소/최대
    assert encode_index([0, 0, 0, 0, 0], [3, 3, 3, 3, 2]) == 0
    assert encode_index([2, 2, 2, 2, 1], [3, 3, 3, 3, 2]) == 161


def test_mixed_radix_bijection_over_full_space():
    cfg = _cfg("_base")
    dims = state_dims(cfg.state)
    assert dims == [3, 3, 3, 3, 2]
    seen = set()
    for combo in itertools.product(*[range(d) for d in dims]):
        idx = encode_index(list(combo), dims)
        seen.add(idx)
    # 162개 조합이 0..161 에 1:1
    assert seen == set(range(cfg.state.state_space_size))


def test_config_injection_changes_classification():
    # revenue=20,000,000 원: _base(th=[10M,50M]) → mid(1), imarketkorea(th=[30M,100M]) → low(0)
    snap = _snapshot(revenue_amount=20_000_000)
    base = build_state(snap, _cfg("_base").state)
    imk = build_state(snap, _cfg("imarketkorea").state)
    assert base.revenue_idx == 1
    assert imk.revenue_idx == 0
    assert base != imk  # 같은 입력이 테넌트 config 에 따라 다른 상태


def test_distribution_unknown_code_raises():
    cfg = _cfg("_base")
    snap = _snapshot(distribution_code="Z")  # code_map 에 없음
    try:
        build_state(snap, cfg.state)
        assert False, "unknown distribution code should raise"
    except ValueError:
        pass


def test_price_zone_and_partner_buckets():
    cfg = _cfg("_base").state
    # 제시가 ≤ 앵커가(9900) → 우선협상 구간(0)
    assert build_state(_snapshot(input_price=9800), cfg).price_zone_idx == 0
    # 제시가 > 앵커가 → 협상 지속 구간(1)
    assert build_state(_snapshot(input_price=10500), cfg).price_zone_idx == 1
    # partner: 0->none(2), 1->single(0), 3->multiple(1)
    assert build_state(_snapshot(partner_count=0), cfg).partner_idx == 2
    assert build_state(_snapshot(partner_count=1), cfg).partner_idx == 0
    assert build_state(_snapshot(partner_count=3), cfg).partner_idx == 1


def test_reward_deterministic_and_config_driven():
    snap = _snapshot(outcome=NegotiationOutcome.FAILURE, round_number=2)
    base = _cfg("_base")
    imk = _cfg("imarketkorea")
    base_rc = RewardCalculator(base.reward, base.state)  # failure_penalty -0.5
    imk_rc = RewardCalculator(imk.reward, imk.state)     # failure_penalty -0.7

    r1 = base_rc.calculate(snap)
    r2 = base_rc.calculate(snap)
    assert r1 == r2                       # 결정론
    assert r1.end_reward == -0.5          # config 반영
    assert imk_rc.calculate(snap).end_reward == -0.7

    # 성공 라운드 보상이 실패보다 크다 (방향성)
    success = _snapshot(outcome=NegotiationOutcome.SUCCESS, round_number=0)
    assert base_rc.calculate(success).total > base_rc.calculate(_snapshot(outcome=NegotiationOutcome.FAILURE, round_number=0)).total


def test_action_card_mapper_roundtrip_and_mask():
    cfg = _cfg("_base")
    mapper = ActionCardMapper(cfg.action_mapping)
    assert mapper.action_space_size == 11
    assert mapper.get_card_id(0) == "NGC-001"
    assert mapper.get_action_id("NGC-001") == 0
    assert mapper.get_card_id(99) is None

    # 중복방지 마스킹: 사용한 action 제외
    mask = mapper.available_mask(used_action_ids={0, 3})
    assert mask.dtype == np.bool_
    assert mask[0] == False and mask[3] == False
    assert mask[1] == True
    assert mask.sum() == 9

    # reload 로 다른 테넌트 카드셋 교체
    mapper.reload(_cfg("imarketkorea").action_mapping)
    assert mapper.get_card_id(0) == "NGC-B001"