feat(cost): persist reasoning tokens and the peak/off-peak tier

ob_ai_call is the only input to V2 pricing, and it was dropping the two
numbers needed to interpret its own cost column.

tokens_reasoning was captured from the provider response but never written to
the table, so the cost of thinking mode was invisible once it left the
process. Folding it into tokens_out was not an option either -- it is billed
at the output rate, and mixing the two hides exactly the difference we now
have to watch.

rate_tier records which price list applied. Without it a row's cost_cents
cannot be rechecked: DeepSeek's peak rate is double off-peak from 2026-08-17,
so recomputing an old row at "the current rate" silently produces a different
answer than what was actually billed.

Migration 0002 adds both columns. tokens_reasoning defaults to 0 rather than
NULL so cost aggregates do not collapse to NULL on historical rows;
rate_tier stays nullable because rows written before this change genuinely
have an unknown tier, and NULL says that more honestly than a guessed 'flat'.

The smoke script now asserts both: reasoning tokens must be 0 (a non-zero
value means a provider default re-enabled thinking mode behind our backs) and
rate_tier must be present. Verified against the real API -- tokens_reasoning=0,
rate_tier=off_peak, 3.9s round trip.

packages/llm/cost.py

@@ -31,6 +31,11 @@ PEAK_WINDOWS: tuple[tuple[time, time], ...] = (
     (time(14, 0), time(18, 0)),
 )
 
+# 记入 ob_ai_call.rate_tier，用于事后复核 cost_cents。
+TIER_PEAK = "peak"
+TIER_OFF_PEAK = "off_peak"
+TIER_FLAT = "flat"  # 只有一套单价（未启用峰谷，或该 provider 不分时段）
+
 
 @dataclass(frozen=True, slots=True)
 class ModelRate:
@@ -63,6 +68,9 @@ class RateTable:
     def known_models(self) -> tuple[str, ...]:
         return tuple(sorted(self._rates))
 
+    def current_tier(self, at: datetime | None = None) -> str:
+        return TIER_FLAT
+
 
 def is_peak(at: datetime | None = None) -> bool:
     """调用时刻是否落在供应商的高峰时段（按北京时间判定）。"""
@@ -94,6 +102,11 @@ class PeakAwareRateTable:
     def known_models(self) -> tuple[str, ...]:
         return self._off_peak.known_models()
 
+    def current_tier(self, at: datetime | None = None) -> str:
+        if not self._enabled:
+            return TIER_FLAT
+        return TIER_PEAK if is_peak(at) else TIER_OFF_PEAK
+
 
 def compute_cost_cents(usage: Usage, rate: ModelRate) -> Decimal:
     """Cost of one call, in cents, with cache-hit tokens billed separately."""
@@ -110,6 +123,10 @@ class RateResolver(Protocol):
 
     def for_model(self, model: str) -> ModelRate: ...
 
+    def current_tier(self) -> str:
+        """此刻适用的档位，随成本一起落库以便事后复核。"""
+        ...
+
 
 @dataclass(frozen=True, slots=True)
 class AICallRecord:
@@ -126,6 +143,7 @@ class AICallRecord:
     user_id: int | None = None
     prompt_id: str | None = None
     prompt_version: str | None = None
+    rate_tier: str | None = None
     error_code: str | None = None
     error_message: str | None = None
     ref_type: str | None = None

packages/llm/gateway.py

@@ -254,6 +254,7 @@ class LLMGateway:
                         user_id=user_id,
                         prompt_id=prompt_id,
                         prompt_version=prompt_version,
+                        rate_tier=self._rates.current_tier(),
                         error_code=exc.code,
                         error_message=str(exc)[:1000],
                         ref_type=ref_type,
@@ -263,6 +264,8 @@ class LLMGateway:
                 raise
 
             latency_ms = self._elapsed_ms(started)
+            # 单价与档位在同一刻取，成本落库后不再回算 —— 跨时段重算会改掉历史成本。
+            rate_tier = self._rates.current_tier()
             cost = compute_cost_cents(response.usage, self._rates.for_model(response.model))
 
             parsed: BaseModel | None = None
@@ -289,6 +292,7 @@ class LLMGateway:
                     user_id=user_id,
                     prompt_id=prompt_id,
                     prompt_version=prompt_version,
+                    rate_tier=rate_tier,
                     error_code=error_code,
                     error_message=last_detail if not success else None,
                     ref_type=ref_type,

packages/llm/recording.py

@@ -36,8 +36,10 @@ class SqlAlchemyAICallRecorder:
                     tokens_in=record.usage.tokens_in,
                     tokens_out=record.usage.tokens_out,
                     tokens_cached=record.usage.tokens_cached,
+                    tokens_reasoning=record.usage.tokens_reasoning,
                     latency_ms=record.latency_ms,
                     cost_cents=record.cost_cents,
+                    rate_tier=record.rate_tier,
                     success=record.success,
                     error_code=record.error_code,
                     error_message=record.error_message,
@@ -59,3 +61,5 @@ class PrometheusAICallRecorder:
         LLM_TOKENS.labels(*labels, "in").inc(record.usage.tokens_in)
         LLM_TOKENS.labels(*labels, "out").inc(record.usage.tokens_out)
         LLM_TOKENS.labels(*labels, "cached").inc(record.usage.tokens_cached)
+        # 思考模式 token 单独出一条曲线：它按输出价计费，是最容易悄悄涨上去的一项。
+        LLM_TOKENS.labels(*labels, "reasoning").inc(record.usage.tokens_reasoning)

packages/shared/models.py

@@ -229,10 +229,16 @@ class AiCall(Base):
     tokens_in: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
     tokens_out: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
     tokens_cached: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    # 思考模式 token 单独记，不并入 tokens_out：它按输出价计费，混在一起就看不出
+    # 开关思考模式的成本差异，而这正是 T-P1-24a 要长期盯的数。
+    tokens_reasoning: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
     latency_ms: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
     cost_cents: Mapped[Decimal] = mapped_column(
         Numeric(14, 6), nullable=False, default=Decimal("0")
     )
+    # 该次调用适用的峰谷档位。没有它就无法事后复核 cost_cents ——
+    # 单价随时刻变化，事后按「当前价」回算会得出不同结果 (T-P1-25c)。
+    rate_tier: Mapped[str | None] = mapped_column(String(16))
     success: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False)
     error_code: Mapped[str | None] = mapped_column(String(64))
     error_message: Mapped[str | None] = mapped_column(Text)