fix(billing): align pre-output Claude refusal settlement with upstream no-charge semantics

[LIghtJUNction]

Upstream

• QuantumNous/new-api issue: [Bug] Claude 上游在输出前 refusal（content 为空、output_tokens 0）时仍按 input_tokens 扣费，Anthropic 对此不计费 QuantumNous/new-api#7551
• QuantumNous/new-api PR: fix(billing): skip billing for Messages API refusals without output QuantumNous/new-api#7552
• Upstream head at review time: e772bed48afaf94b434594e09aeb27fe4ebaea5a
• Reviewed against LMM main@9ce3d78226fb50a38775e6e9d294d5f16771ed59

Problem

Anthropic documents that a refusal which happens before any output is not billed upstream: stop_reason=refusal, empty content, and output_tokens=0. The response may still contain non-zero input-token usage for rate-limit/accounting purposes.

LMM currently detects the refusal in apps/api-go/relay/channel/claude/relay-claude.go, but maybeMarkClaudeRefusal only writes ContextKeyAdminRejectReason. The returned usage then reaches apps/api-go/service/text_quota.go; textQuotaSummary.hasBillableUsage() treats non-zero input tokens as billable, so ratio/fixed/tiered settlement can still charge the user.

This is the same underlying gap fixed by upstream #7552.

Why not directly cherry-pick #7552

Do not mechanically copy the upstream patch yet:

1. #7552 is still open and changes billing policy through a new global opt-in setting.
2. The opt-in exists because Anthropic official endpoints do not charge these refusals, while a third-party Claude-compatible relay may still charge them. Applying a global exemption blindly could create under-billing.
3. LMM has additional billing paths (tiered_expr, quota saturation/composition, tool-call surcharges and custom pre-consume settlement), so the exemption must be placed before all final quota composition paths rather than patched only around the upstream call sites.

Proposed LMM-native implementation

Introduce a narrow settlement marker, e.g. ContextKeyBillingExemptReason=claude_refusal_no_output, only when all of these are true:

• final upstream response is a Claude refusal;
• no content block/output was produced;
• upstream usage exists;
• output_tokens == 0;
• the selected policy says this channel/provider follows Anthropic no-charge refusal semantics.

Keep the original upstream response/usage returned to the client unchanged. Only the billing/logging view should become zero-cost.

For streaming, track whether any content_block_start has occurred before the refusal. For non-streaming, use len(content) == 0. Cover the Claude-native, Chat Completions -> Claude, and Claude -> Responses paths that share the adaptor.

At settlement time, the exemption should be resolved before ordinary ratio billing, fixed-price billing, tiered_expr, and tool-surcharge composition. It should:

• settle final quota to 0;
• fully refund any pre-consumed quota;
• skip tiered expression settlement for this request;
• record request count normally but not token/quota consumption;
• retain the admin-only refusal reason;
• add a machine-readable public/admin log field such as billing_exempt_reason=claude_refusal_no_output;
• never mutate the usage body sent to the client.

Policy decision required

Prefer provider/channel capability over a single unconditional global switch if LMM can distinguish Anthropic-billed channels from third-party Claude-compatible relays. If that distinction is not reliable yet, use an explicit default-off setting as upstream #7552 does.

Do not make client identity, group, /fast, price lock, OAuth mode, or admin-AI state influence this decision.

Acceptance criteria

• Anthropic-style pre-output refusal (content=[], stop_reason=refusal, usage present, output_tokens=0) settles to quota 0 when the no-charge policy is enabled/applicable.
• Pre-consumed quota is fully refunded.
• Ratio billing does not charge input tokens for the exempt request.
• Fixed-price billing does not charge the request.
• tiered_expr does not reintroduce a non-zero charge.
• Tool-surcharge composition cannot reintroduce a charge for an otherwise exempt no-output refusal.
• Mid-stream refusal after any content/output remains normally billable.
• Empty content with output_tokens>0 remains billable.
• Refusal without a usage object remains conservatively billable.
• Non-refusal empty responses are unaffected.
• Policy disabled / unsupported third-party channel preserves current billing behavior.
• Native Claude, converted Chat Completions, streaming, non-streaming, and Claude->Responses paths have regression tests.
• Consume logs distinguish the billing exemption from the existing admin refusal reason.
• Existing pricing/group/model-price-lock//fast/OAuth/admin-AI behavior is unchanged.

Verification target

At minimum run focused Claude relay tests, service settlement tests (including SQLite full pre-consume -> settle flow), tiered-billing regression tests, go vet on touched packages, and the existing CI billing-safety/release qualification jobs.