diff --git a/rust/src/codex_costs/quota_windows/tests.rs b/rust/src/codex_costs/quota_windows/tests.rs
index 004368ac008..fa257ce686b 100644
--- a/rust/src/codex_costs/quota_windows/tests.rs
+++ b/rust/src/codex_costs/quota_windows/tests.rs
@@ -56,6 +56,8 @@ fn cache_with_source_rows(rows: Vec<CodexSourceUsageRow>) -> CostUsageCache {
             mtime_unix_ms: 1,
             prefix_hash: 1,
             rows,
+            scan_since_key: None,
+            scan_until_key: None,
         },
     );
     cache
@@ -101,6 +103,8 @@ fn deduplicates_source_rows_by_file_identity() {
         mtime_unix_ms: 1,
         prefix_hash: 1,
         rows: vec![row(timestamp, 7, 3, Some("gpt-5.5"))],
+        scan_since_key: None,
+        scan_until_key: None,
     };
     let mut cache = cache_with_source_rows(Vec::new());
     cache.codex_source_rows = HashMap::from([
diff --git a/rust/src/core/jsonl_scanner.rs b/rust/src/core/jsonl_scanner.rs
index 3e0a9813bb7..f46a1b69400 100755
--- a/rust/src/core/jsonl_scanner.rs
+++ b/rust/src/core/jsonl_scanner.rs
@@ -303,6 +303,14 @@ pub struct CodexSourceRowCache {
     pub mtime_unix_ms: i64,
     pub prefix_hash: u64,
     pub rows: Vec<CodexSourceUsageRow>,
+    /// Reporting partition the rows were read for. Rows are filtered to that
+    /// partition, so they are recovery evidence only for the same range. Kept
+    /// per entry because a partial pass can replace some entries while the
+    /// cache-wide range keys still describe an earlier completed scan.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub scan_since_key: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub scan_until_key: Option<String>,
 }
 
 /// Per-file usage tracking
diff --git a/rust/src/core/jsonl_scanner/codex.rs b/rust/src/core/jsonl_scanner/codex.rs
index 9a78f8b04da..a839d789730 100644
--- a/rust/src/core/jsonl_scanner/codex.rs
+++ b/rust/src/core/jsonl_scanner/codex.rs
@@ -42,8 +42,9 @@ use crate::cost_reporting_period::cost_bucket_zone;
 /// Codex turn id on source rows so Priority trace evidence can be matched.
 /// Version 7 keys ChatGPT-plan Auto-review requests apart (upstream #4452) so
 /// cached day aggregates from older parsers, which priced them as unpriced,
-/// are rebuilt. Rebuild older artifacts.
-pub(crate) const CODEX_CACHE_SCHEMA_VERSION: u32 = 7;
+/// are rebuilt. Version 8 rebuilds model pricing corrupted by source-row recovery across
+/// different reporting partitions. Rebuild older artifacts.
+pub(crate) const CODEX_CACHE_SCHEMA_VERSION: u32 = 8;
 
 /// Whether a persisted Codex cache artifact matches the current schema.
 /// A mismatched artifact (e.g. a pre-64-bit cache from an older release) is
diff --git a/rust/src/core/jsonl_scanner/codex/source_rows.rs b/rust/src/core/jsonl_scanner/codex/source_rows.rs
index 3d0a8e3a5f5..778cf248561 100644
--- a/rust/src/core/jsonl_scanner/codex/source_rows.rs
+++ b/rust/src/core/jsonl_scanner/codex/source_rows.rs
@@ -183,9 +183,25 @@ pub(crate) fn row_cache(
         mtime_unix_ms: source_mtime_unix_ms(metadata.modified().ok()),
         prefix_hash: source_prefix_hash(file_path, metadata.len())?,
         rows,
+        scan_since_key: None,
+        scan_until_key: None,
     })
 }
 
+impl CodexSourceRowCache {
+    pub(crate) fn record_scan_range(&mut self, range: &CostUsageDayRange) {
+        self.scan_since_key = Some(range.scan_since_key.clone());
+        self.scan_until_key = Some(range.scan_until_key.clone());
+    }
+
+    /// Whether these rows were read for exactly `range`. Entries without a
+    /// recorded range predate per-entry ranges and never match.
+    pub(crate) fn was_read_for(&self, range: &CostUsageDayRange) -> bool {
+        self.scan_since_key.as_deref() == Some(range.scan_since_key.as_str())
+            && self.scan_until_key.as_deref() == Some(range.scan_until_key.as_str())
+    }
+}
+
 /// Verify the retained source prefix before using cached request pricing.
 /// A same-size rewrite is rejected when its modification time changes;
 /// append-only growth is accepted only when the indexed prefix is intact.
diff --git a/rust/src/core/jsonl_scanner/codex/source_rows_tests.rs b/rust/src/core/jsonl_scanner/codex/source_rows_tests.rs
index 1fba1b18f6d..7122e18f860 100644
--- a/rust/src/core/jsonl_scanner/codex/source_rows_tests.rs
+++ b/rust/src/core/jsonl_scanner/codex/source_rows_tests.rs
@@ -147,6 +147,34 @@ fn source_row_cache_rejects_changed_prefix_and_accepts_append() {
     assert!(!row_cache_matches(&path, &changed, &cached));
 }
 
+#[test]
+fn source_row_cache_is_evidence_only_for_the_range_it_was_read_for() {
+    let root = tempfile::tempdir().unwrap();
+    let path = root.path().join("session.jsonl");
+    std::fs::write(&path, b"row\n").unwrap();
+    let metadata = std::fs::metadata(&path).unwrap();
+    let range = |since: &str, until: &str| CostUsageDayRange {
+        since_key: since.to_string(),
+        until_key: until.to_string(),
+        scan_since_key: since.to_string(),
+        scan_until_key: until.to_string(),
+    };
+    let thirty_days = range("2026-09-24", "2026-10-23");
+    let one_day = range("2026-10-23", "2026-10-23");
+
+    let mut cached = row_cache(&path, &metadata, Vec::new()).expect("source cache");
+    assert!(
+        !cached.was_read_for(&thirty_days),
+        "legacy entry has no range"
+    );
+
+    cached.record_scan_range(&one_day);
+    assert!(cached.was_read_for(&one_day));
+    // A shorter pass that replaced this entry must not be read back as
+    // evidence for the longer window, whatever the cache-wide keys say.
+    assert!(!cached.was_read_for(&thirty_days));
+}
+
 #[test]
 fn source_row_cache_hashes_the_entire_cached_prefix() {
     let root = tempfile::tempdir().unwrap();
diff --git a/rust/src/cost_scanner/codex/flat_sources.rs b/rust/src/cost_scanner/codex/flat_sources.rs
index e646e75f64d..eec14a3fd80 100644
--- a/rust/src/cost_scanner/codex/flat_sources.rs
+++ b/rust/src/cost_scanner/codex/flat_sources.rs
@@ -446,6 +446,8 @@ mod tests {
                 mtime_unix_ms: 0,
                 prefix_hash: 0,
                 rows: Vec::new(),
+                scan_since_key: None,
+                scan_until_key: None,
             },
         );
         cache.codex_fork_rows.insert("from".to_string(), Vec::new());
diff --git a/rust/src/cost_scanner/codex/priority_trace/tests/scanner.rs b/rust/src/cost_scanner/codex/priority_trace/tests/scanner.rs
index 5b1fcfb7e49..84dfff7ec64 100644
--- a/rust/src/cost_scanner/codex/priority_trace/tests/scanner.rs
+++ b/rust/src/cost_scanner/codex/priority_trace/tests/scanner.rs
@@ -756,6 +756,8 @@ fn budget_pruning_rebuilds_priority_days_from_retained_files() {
                 mtime_unix_ms: 0,
                 prefix_hash: 0,
                 rows: vec![standard_row(day, input, turn)],
+                scan_since_key: None,
+                scan_until_key: None,
             },
         );
         cursor.request_sources.insert(
diff --git a/rust/src/cost_scanner/codex/scan.rs b/rust/src/cost_scanner/codex/scan.rs
index e6ca018a196..b095aa6377f 100644
--- a/rust/src/cost_scanner/codex/scan.rs
+++ b/rust/src/cost_scanner/codex/scan.rs
@@ -41,6 +41,14 @@ fn codex_source_row_plan(
     let recoverable_cache = cache
         .codex_source_rows
         .get(&path.to_string_lossy().to_string())
+        // Source rows are filtered to the reporting partition. Comparing a
+        // different window mistakes newly included history for lost pricing
+        // evidence and turns otherwise known models into unattributed usage.
+        // A changed partition must establish its evidence from the source;
+        // durable Priority trace evidence is applied separately below. The
+        // range is checked per entry: the cache-wide range keys describe the
+        // last completed scan, and a partial pass can replace entries first.
+        .filter(|cached| cached.was_read_for(scan_range))
         .filter(|cached| !cached.rows.is_empty() && row_cache_matches(path, &metadata, cached))
         .cloned();
     Some(CodexSourceRowPlan {
@@ -53,7 +61,12 @@ fn codex_source_row_plan(
 /// Apply a source-row plan: recover cached pricing when required, retain the
 /// resulting rows as evidence, and unprice the file's day map when recovery
 /// left appended rows unvalidated.
-fn apply_codex_source_row_plan(cache: &mut CostUsageCache, key: &str, plan: CodexSourceRowPlan) {
+fn apply_codex_source_row_plan(
+    cache: &mut CostUsageCache,
+    key: &str,
+    plan: CodexSourceRowPlan,
+    scan_range: &CostUsageDayRange,
+) {
     let CodexSourceRowPlan {
         metadata,
         source_rows,
@@ -68,9 +81,10 @@ fn apply_codex_source_row_plan(cache: &mut CostUsageCache, key: &str, plan: Code
     } else {
         source_rows
     };
-    let Some(source_cache) = row_cache(&key_path(key), &metadata, rows) else {
+    let Some(mut source_cache) = row_cache(&key_path(key), &metadata, rows) else {
         return;
     };
+    source_cache.record_scan_range(scan_range);
     if requires_recovery && let Some(usage) = cache.files.get_mut(key) {
         // Rebuild from recovered rows even when a new row is intentionally
         // unresolved. Leaving the normal parser's model-priced days in place
@@ -413,7 +427,7 @@ pub(super) fn scan_codex_detailed_with_cache(
             incomplete_processed.push(key);
             stats.files_deferred = stats.files_deferred.saturating_add(1);
         } else if let Some(plan) = codex_source_row_plan(&cache, &candidate.path, scan_range) {
-            apply_codex_source_row_plan(&mut cache, &key, plan);
+            apply_codex_source_row_plan(&mut cache, &key, plan, scan_range);
         }
     }
     unprocessed.extend(deferred_paths);
diff --git a/rust/src/cost_scanner/tests/period.rs b/rust/src/cost_scanner/tests/period.rs
index 1daf48ff68b..de4f32b9f8a 100644
--- a/rust/src/cost_scanner/tests/period.rs
+++ b/rust/src/cost_scanner/tests/period.rs
@@ -153,3 +153,59 @@ fn codex_month_to_date_excludes_the_previous_month() {
     );
     assert_eq!(mtd, 100);
 }
+
+#[test]
+fn changing_history_windows_keeps_source_model_pricing_complete() {
+    let root = tempfile::tempdir().unwrap();
+    let sessions = root.path().join("sessions");
+    let cache = root.path().join("cache");
+    let today = Local::now().date_naive();
+    let folder = sessions.join(today.format("%Y/%m/%d").to_string());
+    std::fs::create_dir_all(&folder).unwrap();
+    let mut lines = Vec::new();
+    for age in [8, 1, 0] {
+        let timestamp = Local
+            .from_local_datetime(&(today - Duration::days(age)).and_hms_opt(12, 0, 0).unwrap())
+            .earliest()
+            .unwrap()
+            .with_timezone(&Utc)
+            .to_rfc3339();
+        lines.push(
+            serde_json::json!({
+                "type": "turn_context", "timestamp": timestamp,
+                "payload": {"model": "gpt-5"}
+            })
+            .to_string(),
+        );
+        lines.push(serde_json::json!({
+            "type": "event_msg", "timestamp": timestamp,
+            "payload": {"type": "token_count", "info": {
+                "last_token_usage": {"input_tokens": 100, "cached_input_tokens": 0, "output_tokens": 0}
+            }}
+        }).to_string());
+    }
+    std::fs::write(folder.join("history.jsonl"), lines.join("\n") + "\n").unwrap();
+    for days in [30, 1, 7, 30, 1, 30] {
+        let summary = CostScanner::new(days)
+            .with_options(CostScanOptions::app_driven())
+            .with_cache_root(&cache)
+            .with_sessions_dirs(vec![sessions.clone()])
+            .scan_codex();
+        assert!(summary.history_coverage_established);
+        assert!(
+            !summary.model_pricing_completeness.is_partial(),
+            "known source models lost pricing after switching to {days} days: {:?}",
+            summary.model_pricing_completeness
+        );
+        assert_eq!(
+            summary.input_tokens,
+            if days >= 9 {
+                300
+            } else if days > 1 {
+                200
+            } else {
+                100
+            }
+        );
+    }
+}