fix(llm): carry the cache-write TTL split from Anthropic message_start usage

crates/llm/src/conversion/messages.rs

@@ -801,10 +801,22 @@ pub mod from_completions {
 								message.usage.cache_read_input_tokens.map(|i| i as u64);
 							r.response.cache_creation_input_tokens =
 								message.usage.cache_creation_input_tokens.map(|i| i as u64);
-							// TTL split stays unread here — message_start usage
-							// is provisional pre-turn state; only the terminal
-							// message_delta is cumulative-final (see the
-							// passthrough MessageStart arm).
+							// Input-side usage (cache write split included) is final at
+							// `message_start`; a reported all-zero split with a
+							// small aggregate is the provider's authoritative
+							// sub-minimum no-op write evidence.
+							r.response.cache_creation_5m_input_tokens = message
+								.usage
+								.cache_creation
+								.as_ref()
+								.and_then(|c| c.ephemeral_5m_input_tokens)
+								.map(|i| i as u64);
+							r.response.cache_creation_1h_input_tokens = message
+								.usage
+								.cache_creation
+								.as_ref()
+								.and_then(|c| c.ephemeral_1h_input_tokens)
+								.map(|i| i as u64);
 							// Provisional evidence: the cumulative totals only
 							// arrive on the terminal `message_delta`.
 							r.response.usage_complete = Some(false);
@@ -946,11 +958,19 @@ pub mod from_completions {
 						if let Some(cwt) = usage.cache_creation_input_tokens {
 							r.response.cache_creation_input_tokens = Some(cwt as u64);
 						}
-						if let Some(split) = &usage.cache_creation {
-							r.response.cache_creation_5m_input_tokens =
-								split.ephemeral_5m_input_tokens.map(|i| i as u64);
-							r.response.cache_creation_1h_input_tokens =
-								split.ephemeral_1h_input_tokens.map(|i| i as u64);
+						// The cumulative delta wins when it carries the split;
+						// absent leaves earlier evidence untouched.
+						if usage.cache_creation.is_some() {
+							r.response.cache_creation_5m_input_tokens = usage
+								.cache_creation
+								.as_ref()
+								.and_then(|c| c.ephemeral_5m_input_tokens)
+								.map(|i| i as u64);
+							r.response.cache_creation_1h_input_tokens = usage
+								.cache_creation
+								.as_ref()
+								.and_then(|c| c.ephemeral_1h_input_tokens)
+								.map(|i| i as u64);
 						}
 						if let Some(o) = usage.output_tokens {
 							r.response.output_tokens = Some(o as u64);
@@ -1195,13 +1215,26 @@ pub fn passthrough_stream(
 							message.usage.cache_read_input_tokens.map(|i| i as u64);
 						r.response.cache_creation_input_tokens =
 							message.usage.cache_creation_input_tokens.map(|i| i as u64);
-						// NOTE: the TTL split is deliberately NOT read here.
-						// `message_start` usage is provisional pre-turn state —
-						// captured production evidence shows it reporting
-						// `cache_creation: {5m: 0, 1h: 0}` alongside a positive
-						// aggregate — while the terminal `message_delta` usage
-						// is cumulative-final. Only the delta may carry the
-						// split into evidence.
+						// Input-side usage (cache write split included) is
+						// final at `message_start`; the terminal
+						// `message_delta` is cumulative-final. Only
+						// self-consistent splits become evidence.
+						// Input-side usage (cache write split included) is final at
+						// `message_start`; a reported all-zero split with a small
+						// aggregate is the provider's authoritative sub-minimum
+						// no-op write evidence.
+						r.response.cache_creation_5m_input_tokens = message
+							.usage
+							.cache_creation
+							.as_ref()
+							.and_then(|c| c.ephemeral_5m_input_tokens)
+							.map(|i| i as u64);
+						r.response.cache_creation_1h_input_tokens = message
+							.usage
+							.cache_creation
+							.as_ref()
+							.and_then(|c| c.ephemeral_1h_input_tokens)
+							.map(|i| i as u64);
 						// Provisional evidence: the cumulative totals only
 						// arrive on the terminal `message_delta`.
 						r.response.usage_complete = Some(false);
@@ -1265,11 +1298,19 @@ pub fn passthrough_stream(
 					if let Some(cwt) = usage.cache_creation_input_tokens {
 						r.response.cache_creation_input_tokens = Some(cwt as u64);
 					}
-					if let Some(split) = &usage.cache_creation {
-						r.response.cache_creation_5m_input_tokens =
-							split.ephemeral_5m_input_tokens.map(|i| i as u64);
-						r.response.cache_creation_1h_input_tokens =
-							split.ephemeral_1h_input_tokens.map(|i| i as u64);
+					// The cumulative delta wins when it carries the split;
+					// absent leaves earlier evidence untouched.
+					if usage.cache_creation.is_some() {
+						r.response.cache_creation_5m_input_tokens = usage
+							.cache_creation
+							.as_ref()
+							.and_then(|c| c.ephemeral_5m_input_tokens)
+							.map(|i| i as u64);
+						r.response.cache_creation_1h_input_tokens = usage
+							.cache_creation
+							.as_ref()
+							.and_then(|c| c.ephemeral_1h_input_tokens)
+							.map(|i| i as u64);
 					}
 					if let Some(o) = usage.output_tokens {
 						r.response.output_tokens = Some(o as u64);

crates/llm/src/conversion/messages_tests.rs

@@ -492,3 +492,43 @@ async fn passthrough_stream_cache_usage_is_preserved() {
 		"cache_creation evidence must survive extraction"
 	);
 }
+
+/// Captured production `message_start` shape (golden stream_basic): a
+/// small aggregate alongside an all-zero split — the provider's
+/// authoritative sub-minimum no-op write evidence. The split buckets are
+/// forwarded EXACTLY as reported (Some(0)/Some(0) — real evidence, not
+/// absence) and the terminal delta, which carries no split here, leaves
+/// that evidence untouched instead of clobbering it.
+#[tokio::test]
+async fn passthrough_stream_sub_minimum_zero_split_is_authoritative_evidence() {
+	let input = format!(
+		"{}\n\n{}\n\n",
+		message_start_with_usage(
+			r#"{"input_tokens":15,"output_tokens":1,"cache_creation_input_tokens":13,"cache_read_input_tokens":12,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0}}"#
+		),
+		r#"data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"output_tokens":21}}"#,
+	);
+	let info = run_passthrough(input).await;
+	assert_eq!(info.response.cache_creation_input_tokens, Some(13));
+	assert_eq!(info.response.cache_creation_5m_input_tokens, Some(0));
+	assert_eq!(info.response.cache_creation_1h_input_tokens, Some(0));
+}
+
+/// Real Anthropic streaming shape (production canary): the input-side
+/// usage is final at `message_start` and SELF-CONSISTENT
+/// (aggregate == 5m + 1h), while the terminal delta carries no split. The
+/// consistent split becomes evidence at start and survives the delta.
+#[tokio::test]
+async fn passthrough_stream_consistent_message_start_split_becomes_evidence() {
+	let input = format!(
+		"{}\n\n{}\n\n",
+		message_start_with_usage(
+			r#"{"input_tokens":10,"output_tokens":1,"cache_creation_input_tokens":67759,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":67759,"ephemeral_1h_input_tokens":0}}"#
+		),
+		r#"data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"output_tokens":44}}"#,
+	);
+	let info = run_passthrough(input).await;
+	assert_eq!(info.response.cache_creation_input_tokens, Some(67759));
+	assert_eq!(info.response.cache_creation_5m_input_tokens, Some(67759));
+	assert_eq!(info.response.cache_creation_1h_input_tokens, Some(0));
+}

crates/llm/src/tests/response/anthropic/stream_basic.messages-completions-streaming.snap

@@ -1,5 +1,6 @@
 ---
 source: crates/llm/src/golden_tests.rs
+assertion_line: 846
 description: src/tests/response/anthropic/stream_basic.json
 ---
 data: {"id":"msg_016wbjv5k86nxxd8CEZwSQnA","choices":[{"index":0,"delta":{"content":"Hi there! How are","role":"assistant"}}],"created":123,"model":"claude-haiku-4-5-20251001","service_tier":"standard","system_fingerprint":null,"object":"chat.completion.chunk","usage":null}
@@ -23,6 +24,8 @@ data: [DONE]
   "output_tokens": 21,
   "total_tokens": 36,
   "cache_creation_input_tokens": 13,
+  "cache_creation_5m_input_tokens": 0,
+  "cache_creation_1h_input_tokens": 0,
   "cached_input_tokens": 12,
   "service_tier": "standard",
   "provider_model": "claude-haiku-4-5-20251001",

crates/llm/src/tests/response/anthropic/stream_basic.messages-messages-streaming.snap

@@ -1,5 +1,6 @@
 ---
 source: crates/llm/src/golden_tests.rs
+assertion_line: 846
 description: src/tests/response/anthropic/stream_basic.json
 ---
 event: message_start
@@ -42,6 +43,8 @@ data: {"type":"message_stop"            }
   "output_tokens": 21,
   "total_tokens": 36,
   "cache_creation_input_tokens": 13,
+  "cache_creation_5m_input_tokens": 0,
+  "cache_creation_1h_input_tokens": 0,
   "cached_input_tokens": 12,
   "service_tier": "standard",
   "provider_model": "claude-haiku-4-5-20251001",

crates/llm/src/tests/response/anthropic/stream_thinking.messages-completions-streaming.snap

@@ -1,5 +1,6 @@
 ---
 source: crates/llm/src/golden_tests.rs
+assertion_line: 846
 description: src/tests/response/anthropic/stream_thinking.json
 ---
 data: {"id":"msg_018XasA6jgabsibaUtstx4bg","choices":[{"index":0,"delta":{"content":null,"role":"assistant","reasoning_content":"The user is"}}],"created":123,"model":"claude-opus-4-6","service_tier":"standard","system_fingerprint":null,"object":"chat.completion.chunk","usage":null}
@@ -31,6 +32,8 @@ data: [DONE]
   "output_tokens": 264,
   "total_tokens": 311,
   "cache_creation_input_tokens": 0,
+  "cache_creation_5m_input_tokens": 0,
+  "cache_creation_1h_input_tokens": 0,
   "cached_input_tokens": 0,
   "service_tier": "standard",
   "provider_model": "claude-opus-4-6",

crates/llm/src/tests/response/anthropic/stream_thinking.messages-messages-streaming.snap

@@ -1,5 +1,6 @@
 ---
 source: crates/llm/src/golden_tests.rs
+assertion_line: 846
 description: src/tests/response/anthropic/stream_thinking.json
 ---
 event: message_start
@@ -60,6 +61,8 @@ data: {"type":"message_stop"          }
   "output_tokens": 264,
   "total_tokens": 311,
   "cache_creation_input_tokens": 0,
+  "cache_creation_5m_input_tokens": 0,
+  "cache_creation_1h_input_tokens": 0,
   "cached_input_tokens": 0,
   "service_tier": "standard",
   "provider_model": "claude-opus-4-6",

crates/llm/src/types/messages.rs

@@ -169,6 +169,11 @@ impl Usage {
 	/// wire format stays byte-identical — no re-parsed typed field, no key
 	/// reordering. `None` = the provider reported no split; consumers must
 	/// not guess the class from the aggregate.
+	///
+	/// A reported all-zero split next to a small aggregate is the
+	/// provider's authoritative sub-minimum no-op write evidence (observed
+	/// on captured production usage): the classes are authoritative and the
+	/// residual aggregate meters at zero — not an inconsistency to hide.
 	pub fn cache_ttl_split(&self) -> (Option<u64>, Option<u64>) {
 		let cache_creation = self.rest.get("cache_creation");
 		let five_m = cache_creation