diff --git a/.agents/docs/performance-strategy.md b/.agents/docs/performance-strategy.md
index 5f96e0f..dffa70a 100644
--- a/.agents/docs/performance-strategy.md
+++ b/.agents/docs/performance-strategy.md
@@ -164,12 +164,16 @@ This checkpoint is historical and predates the breaking design: the final API de
 7. Consider de-monomorphizing `relative` for consumer code size. Treat this as compile-time/binary-size work unless runtime evidence says otherwise.
 8. Attempt explicit SIMD only if instruction and profile data still show separator or common-prefix scanning as a leading cost. Completed for the remaining case-sensitive common-prefix loop by `0da5a61`; existing `memchr` scans and Windows comparison stay unchanged.
 
-## Remaining consumer-backed directions
+## Remaining library directions (corrected)
 
-- Keep the Windows native-separator-aware scan under the native CI allocation and correctness gates. The current snapshots prove a zero-allocation canonical descendant and one-allocation ordinary final `String`; the drive, UNC, verbatim, device, mixed-separator, invalid-wide, and exact-spelling matrix must remain intact when the scan changes.
-- Migrate Rolldown's known-UTF-8 `stabilize_id` and sourcemap endpoints to `relative(base).into_owned().into_slash()` without changing their distinct empty-result handling. The target is one reusable final `String` buffer with no second result allocation. Do not use a lossy-named operation, and reconsider a public fused method only if an end-to-end final-container benchmark later shows that ordinary composition cannot meet the target.
-- Consider making normalized module IDs an explicit Rolldown invariant. All 4,901 observed normalization calls were clean borrowed no-ops, so validating once at an ingestion boundary could remove repeated scans more effectively than another SugarPath classifier branch. This is a Rolldown data-model change, not a library fast path.
-- Revisit direct ArcStr initialization only with a caller-owned output sink or exact-length writer benchmark. ArcStr cannot adopt the current `String` allocation, and a SugarPath dependency on ArcStr would be inappropriate; the integration boundary must stay generic and prove that its callback or writer overhead is lower than the saved copy.
+Clean-path fast cases (0-alloc normalize, absolute absolutize, borrowable descendant `relative`, clean owned normalize/slash, owned-cwd `absolutize_with` reuse) are done and gated. **Slow path means the full general algorithm, not a rare edge:** dirty normalize, non-borrowable `relative` (upward, unequal parents, complex roots), dirty join pipelines, and invalid-encoding rebuilds. Those are the default path for many inputs and remain the **in-crate optimization target**.
+
+- **Recorded unequal leading-parent relative improvement:** pure relative pairs that miss the equal-parent lexical hit now resolve both sides against one ambient cwd as normal-component stacks and allocate only the result buffer (allocation hard gate: 5 alloc / multi-realloc → 1 / 0 for shallow and p99 unequal rows). Absolute/absolute and equal-parent hits still return earlier and are untaxed. Same-machine Criterion on Apple ARM64 with `cached_current_dir` saw about 50% wall-time reduction on the shallow unequal workload and about 35% on the p99 unequal workload (directional local evidence; continuous CodSpeed remains Linux corroboration).
+- Still open on the general path: dirty join→PathBuf/String intermediate costs (2–3 alloc), invalid-encoding absolute relative rebuilds, and any further cheapening of dirty `normalize_inner` without regressing clean 0-alloc gates.
+- Keep the Windows native-separator-aware scan under native CI allocation and correctness gates (drive, UNC, verbatim, device, mixed-separator, invalid-wide, exact spelling).
+- De-monomorphize `relative` only with binary-size / `llvm-lines` evidence. More SIMD only with profile proof. No public fused relative-to-string API and no ArcStr dependency in this crate by default.
+
+Consumer adoption and app-level invariants are **out of this library backlog**.
 
 ## Boshen plan applicability audit
 
diff --git a/benchmarks/allocations/x86_64-pc-windows-msvc.snap b/benchmarks/allocations/x86_64-pc-windows-msvc.snap
index 8fe02f1..1fa9963 100644
--- a/benchmarks/allocations/x86_64-pc-windows-msvc.snap
+++ b/benchmarks/allocations/x86_64-pc-windows-msvc.snap
@@ -48,9 +48,9 @@ Each row measures one operation after allocation-capable setup and an untracked
 | relative / relative inputs -> relative path | 1 | 0 | 0 |
 | relative / dotted relative inputs -> relative path | 1 | 0 | 0 |
 | relative / equal leading-parent inputs -> relative path | 1 | 0 | 0 |
-| relative / unequal leading-parent inputs -> relative path | 5 | 0 | 2 |
+| relative / unequal leading-parent inputs -> relative path | 1 | 0 | 0 |
 | relative / p99-depth equal leading-parent inputs -> relative path | 1 | 0 | 0 |
-| relative / p99-depth unequal leading-parent inputs -> relative path | 5 | 0 | 2 |
+| relative / p99-depth unequal leading-parent inputs -> relative path | 1 | 0 | 0 |
 | relative / current-directory spellings -> relative path | 0 | 0 | 0 |
 | relative / dotted slow path -> relative path | 4 | 0 | 0 |
 | relative / invalid-encoding absolute target -> relative path (setup excluded) | 3 | 0 | 0 |
diff --git a/benchmarks/allocations/x86_64-unknown-linux-gnu.snap b/benchmarks/allocations/x86_64-unknown-linux-gnu.snap
index 849e941..24c8c9b 100644
--- a/benchmarks/allocations/x86_64-unknown-linux-gnu.snap
+++ b/benchmarks/allocations/x86_64-unknown-linux-gnu.snap
@@ -45,9 +45,9 @@ Each row measures one operation after allocation-capable setup and an untracked
 | relative / relative inputs -> relative path | 1 | 0 | 0 |
 | relative / dotted relative inputs -> relative path | 1 | 0 | 0 |
 | relative / equal leading-parent inputs -> relative path | 1 | 0 | 0 |
-| relative / unequal leading-parent inputs -> relative path | 5 | 0 | 4 |
+| relative / unequal leading-parent inputs -> relative path | 1 | 0 | 0 |
 | relative / p99-depth equal leading-parent inputs -> relative path | 1 | 0 | 0 |
-| relative / p99-depth unequal leading-parent inputs -> relative path | 5 | 0 | 6 |
+| relative / p99-depth unequal leading-parent inputs -> relative path | 1 | 0 | 0 |
 | relative / current-directory spellings -> relative path | 0 | 0 | 0 |
 | relative / dotted slow path -> relative path | 1 | 0 | 0 |
 | relative / invalid-encoding absolute target -> relative path (setup excluded) | 3 | 0 | 3 |
diff --git a/src/impl_sugar_path.rs b/src/impl_sugar_path.rs
index 0944c39..af00b48 100644
--- a/src/impl_sugar_path.rs
+++ b/src/impl_sugar_path.rs
@@ -118,18 +118,61 @@ fn try_relative_lexically(target: &Path, base: &Path) -> Option<PathBuf> {
 
   let target = collect_lexical_normals(target, target_shape);
   let base = collect_lexical_normals(base, base_shape);
+  relative_from_normal_stacks(&base, &target)
+}
+
+/// Walk an absolute path into its normal-component stack under the root (and
+/// Windows prefix). CurDir is ignored; ParentDir pops; RootDir clears normals.
+/// Pure relative inputs share this stack when resolved against one cwd.
+fn absolute_normal_stack<'a>(path: &'a Path) -> Option<OsStrVec<'a>> {
+  if !path.is_absolute() {
+    return None;
+  }
 
-  let common_len = target
+  let mut stack = OsStrVec::with_capacity(16);
+  for component in path.components() {
+    match component {
+      Component::Prefix(_) => {
+        // Pure relative inputs inherit the cwd prefix implicitly. Different
+        // prefixes cannot arise when both sides resolve against the same cwd.
+      }
+      Component::RootDir => stack.clear(),
+      Component::CurDir => {}
+      Component::ParentDir => {
+        let _ = stack.pop();
+      }
+      Component::Normal(normal) => stack.push(normal),
+    }
+  }
+  Some(stack)
+}
+
+fn apply_relative_shape_to_stack<'a>(
+  cwd_stack: &OsStrVec<'a>,
+  shape: LexicalRelativeShape,
+  relative: &'a Path,
+) -> OsStrVec<'a> {
+  let kept = cwd_stack.len().saturating_sub(shape.unresolved_parents);
+  let mut stack = OsStrVec::with_capacity(kept + shape.surviving_normals);
+  stack.extend_from_slice(&cwd_stack[..kept]);
+  for normal in collect_lexical_normals(relative, shape) {
+    stack.push(normal);
+  }
+  stack
+}
+
+fn relative_from_normal_stacks(base: &OsStrVec<'_>, target: &OsStrVec<'_>) -> Option<PathBuf> {
+  let common_len = base
     .iter()
-    .zip(&base)
-    .take_while(|(target, base)| {
+    .zip(target.iter())
+    .take_while(|(from, to)| {
       #[cfg(target_family = "windows")]
       {
-        target.eq_ignore_ascii_case(base)
+        from.eq_ignore_ascii_case(to)
       }
       #[cfg(not(target_family = "windows"))]
       {
-        target == base
+        from == to
       }
     })
     .count();
@@ -137,13 +180,15 @@ fn try_relative_lexically(target: &Path, base: &Path) -> Option<PathBuf> {
   let up_len = base.len() - common_len;
   let target_suffix = &target[common_len..];
   #[cfg(target_family = "windows")]
-  if up_len == 0
-    && target_suffix.first().is_some_and(|component| {
-      !windows_standalone_relative_bytes_are_representable(component.as_encoded_bytes())
-    })
+  if target_suffix.iter().any(|component| memchr(b'/', component.as_encoded_bytes()).is_some())
+    || (up_len == 0
+      && target_suffix.first().is_some_and(|component| {
+        !windows_standalone_relative_bytes_are_representable(component.as_encoded_bytes())
+      }))
   {
     return None;
   }
+
   let component_count = up_len + target_suffix.len();
   let capacity = up_len * 2
     + target_suffix.iter().map(|component| component.len()).sum::<usize>()
@@ -160,17 +205,42 @@ fn try_relative_lexically(target: &Path, base: &Path) -> Option<PathBuf> {
     Some(PathBuf::from(relative))
   }
   #[cfg(not(target_family = "windows"))]
-  let mut relative = PathBuf::with_capacity(capacity);
-  #[cfg(not(target_family = "windows"))]
-  for _ in 0..up_len {
-    relative.push(Component::ParentDir);
-  }
-  #[cfg(not(target_family = "windows"))]
-  for component in target_suffix {
-    relative.push(component);
+  {
+    let mut relative = PathBuf::with_capacity(capacity);
+    for _ in 0..up_len {
+      relative.push(Component::ParentDir);
+    }
+    for component in target_suffix {
+      relative.push(*component);
+    }
+    Some(relative)
   }
-  #[cfg(not(target_family = "windows"))]
-  Some(relative)
+}
+
+/// When both inputs are pure relative but do not share a leading-parent count
+/// (or other cwd-independent shape), resolve them against one absolute cwd as
+/// normal-component stacks and build a single relative result. Avoids cloning
+/// cwd twice and allocating two intermediate absolute `PathBuf`s.
+///
+/// Callers that already hold shapes should use
+/// [`relative_both_relative_via_cwd_with_shapes`] so each path is classified once.
+fn try_relative_both_relative_via_cwd(target: &Path, base: &Path, cwd: &Path) -> Option<PathBuf> {
+  let target_shape = classify_lexical_relative(target)?;
+  let base_shape = classify_lexical_relative(base)?;
+  relative_both_relative_via_cwd_with_shapes(target, target_shape, base, base_shape, cwd)
+}
+
+fn relative_both_relative_via_cwd_with_shapes(
+  target: &Path,
+  target_shape: LexicalRelativeShape,
+  base: &Path,
+  base_shape: LexicalRelativeShape,
+  cwd: &Path,
+) -> Option<PathBuf> {
+  let cwd_stack = absolute_normal_stack(cwd)?;
+  let base_resolved = apply_relative_shape_to_stack(&cwd_stack, base_shape, base);
+  let target_resolved = apply_relative_shape_to_stack(&cwd_stack, target_shape, target);
+  relative_from_normal_stacks(&base_resolved, &target_resolved)
 }
 
 #[cfg(target_family = "windows")]
@@ -823,7 +893,30 @@ fn try_relative_outcome<'a>(
     return Ok(outcome);
   }
 
-  // Slow path: avoid current_dir() for already-absolute paths.
+  // Pure lexical relative pairs only (no prefix/root). Windows drive-relative
+  // inputs are `!has_root()` but carry a Prefix — they must keep try_absolutize
+  // so ambient relative uses per-drive `std::path::absolute`, not absolutize_with
+  // against the process cwd. Classify once; reuse shapes for the stack resolve.
+  if let (Some(target_shape), Some(base_shape)) =
+    (classify_lexical_relative(target_path), classify_lexical_relative(base_path))
+  {
+    let cwd = try_get_current_dir()?;
+    if let Some(relative) = relative_both_relative_via_cwd_with_shapes(
+      target_path,
+      target_shape,
+      base_path,
+      base_shape,
+      cwd.as_ref(),
+    ) {
+      return Ok(RelativeOutcome::Native(relative));
+    }
+    let base = base_path.absolutize_with(cwd.as_ref()).into_owned();
+    let target = target_path.absolutize_with(cwd.as_ref()).into_owned();
+    return Ok(relative_from_resolved(base, target));
+  }
+
+  // Slow path: avoid current_dir() for already-absolute paths. Windows
+  // drive-relative receivers still take try_absolutize here.
   let base = if base_path.is_absolute() {
     normalize_for_resolution(base_path).into_owned()
   } else {
@@ -852,6 +945,13 @@ where
 
   assert!(cwd.as_ref().is_absolute(), "explicit current directory must be absolute");
 
+  // Same pure-lexical gate as ambient relative. try_relative_both_relative_via_cwd
+  // already requires classification; do not treat every !has_root path as pure
+  // relative (Windows drive-relative keeps absolutize_with / drive rules below).
+  if let Some(relative) = try_relative_both_relative_via_cwd(target_path, base_path, cwd.as_ref()) {
+    return RelativeOutcome::Native(relative);
+  }
+
   let base = if base_path.is_absolute() {
     normalize_for_resolution(base_path).into_owned()
   } else {
diff --git a/tests/relative_lexical.rs b/tests/relative_lexical.rs
index d0436a9..f0de3a4 100644
--- a/tests/relative_lexical.rs
+++ b/tests/relative_lexical.rs
@@ -148,6 +148,32 @@ fn deep_unequal_parent_miss_uses_the_cwd_dependent_fallback() {
   assert_eq!(target.relative(base).as_os_str(), expected.as_os_str());
 }
 
+/// Ambient `relative` for Windows drive-relative inputs must use `try_absolutize`
+/// (per-drive `std::path::absolute`), not the pure-lexical shared-cwd stack path.
+/// Drive-relative paths are `!has_root()` but carry a Prefix — treating every
+/// `!has_root` pair as pure relative regressed the different-drive allocation row.
+#[cfg(windows)]
+#[test]
+fn ambient_drive_relative_relative_matches_try_absolutize_composition() {
+  for (target, base) in [
+    (r"C:dist\assets\index.js", r"C:dist\chunks"),
+    (r"D:dist\assets\index.js", r"C:dist\chunks"),
+    (r"C:dist\assets\index.js", r"D:dist\chunks"),
+  ] {
+    let target_path = Path::new(target);
+    let base_path = Path::new(base);
+    let actual = target_path.relative(base_path);
+    let resolved_base = base_path.try_absolutize().expect("absolutize base").into_owned();
+    let resolved_target = target_path.try_absolutize().expect("absolutize target").into_owned();
+    let expected = resolved_target.relative(resolved_base.as_path());
+    assert_eq!(
+      actual.as_os_str(),
+      expected.as_os_str(),
+      "target {target:?}, base {base:?}: ambient relative must match try_absolutize composition",
+    );
+  }
+}
+
 #[cfg(windows)]
 #[test]
 fn windows_explicit_cwd_preserves_or_cancels_drive_relative_context() {