fix(billing): settle tiered retries with final group (#6518)

pkg/billingexpr/expr.md

@@ -159,7 +159,7 @@ When a request arrives and the model uses `tiered_expr` billing:
 2. Builds `RequestInput` (headers + body) for `param()` / `header()` functions
 3. Runs expression with estimated tokens: `RunExprWithRequest(expr, {P, C}, requestInput)`
 4. Converts output to quota: `rawCost / 1,000,000 * QuotaPerUnit`
-5. Creates `BillingSnapshot` (frozen state for settlement) and stores on `RelayInfo`
+5. Creates `BillingSnapshot` and stores it on `RelayInfo`. Expression and request state stay frozen for settlement. An auto-group retry refreshes group-dependent fields from the selected group before the next upstream attempt. If a free initial group skipped pre-consume and the retry selects a paid group, the billing session is created before that attempt. If an existing session moves to a more expensive group, its reservation is raised to that group's estimate before sending; cheaper groups are refunded only after actual usage is settled.
 
 ### 4. Settlement (Actual Billing)
 
@@ -173,7 +173,7 @@ After the upstream response returns with actual token usage:
    - For Claude-format APIs (input_tokens is text-only): no adjustment needed
 
 2. `TryTieredSettle(relayInfo, params)`:
-   - Uses the frozen `BillingSnapshot` from pre-consume
+   - Uses the captured `BillingSnapshot`, whose group-dependent fields have been refreshed from the final selected group
    - Re-runs the expression with actual token counts
    - Converts via `quotaConversion()` (version-dispatched)
    - Returns actual quota

pkg/billingexpr/types.go

@@ -36,8 +36,10 @@ type TraceResult struct {
 	Cost        float64 `json:"cost"`
 }
 
-// BillingSnapshot captures the billing rule state frozen at pre-consume time.
-// It is fully serializable and contains no compiled program pointers.
+// BillingSnapshot captures billing state at pre-consume time. Expression and
+// request fields stay frozen; group-dependent fields are refreshed before an
+// auto-group retry and settlement. It is fully serializable and contains no
+// compiled program pointers.
 type BillingSnapshot struct {
 	BillingMode               string  `json:"billing_mode"`
 	ModelName                 string  `json:"model_name"`