perf: reduce redundant scans in path hot paths

.agents/docs/performance-strategy.md

@@ -24,6 +24,12 @@ The 20 pinned Rolldown `normalize` call sites also show why the next owner matte
 
 If a real Rolldown run is instrumented, keep aggregate-only counters outside timed runs: `borrowed_clean`, `dirty_early`, `dirty_late`, `invalid_encoding`, total input bytes, and bytes examined before the decision. Separate hot module-ID normalization, the relative pipeline, and other call sites; never record raw paths, hashes, or reversible samples.
 
+The current safe normalization classifier reads `OsStr` encoded bytes directly for ASCII dot and native-separator checks on conventional Unix targets and Windows; Cygwin retains its existing Unicode-validation gate. This lets clean non-UTF-8 native encoding borrow without rebuilding on the encoded-byte targets, while exhaustive short-path tests compare classifier dispatch with unconditional full normalization for arbitrary Unix bytes and Windows wide units. The counterweight is non-UTF-8 encoding combined with dirty spelling: the former Unicode gate returned immediately, whereas the encoded-byte classifier must continue to the dirty marker. In the stacked Linux CodSpeed comparison, `non_utf8_dirty_before` and `non_utf8_before_dirty_late` were each 16.27% slower, while `non_utf8_lexically_clean` was 2.4 times as efficient and `non_utf8_absolute_target` was 17.33% more efficient with peak memory reduced from 149 to 64 bytes. Keep these rows separate rather than averaging a rare non-UTF-8-and-dirty cost into the observed clean valid Rolldown workload.
+
+The explicitly renamed `borrowed_receiver/non_utf8_strict` row reports an additional 11.68% slowdown. Its Unix implementation remains the same `Path::to_str().map(Cow::Borrowed)` path between the stacked base and head, so this result reflects whole-crate compiled layout rather than added slash-conversion work. The vouched UTF-8 performance target keeps all three non-UTF-8 regressions visible without treating them as acceptance gates.
+
+Generated-code inspection supports outlining `needs_normalization` while leaving the leading-parent helper and `normalize_inner` eligible for inlining. Linux CodSpeed showed broad classifier gains with the classifier outlined, but forcing `normalize_inner` out of line regressed the existing parent-cancellation benchmark by 12.82%; dirty normalization remains a supported general path rather than cold code.
+
 ## Recorded Rolldown runtime shape trace
 
 Aggregate-only instrumentation used SugarPath commit `60b74dc` and Rolldown commit [`b9823050b`](https://github.com/rolldown/rolldown/commit/b9823050bc658ef65105148ea0504d4fbda7fa4c). Ten full ThreeJS generate/write/sourcemap builds produced 3,696 normalization calls, and one Rome TypeScript build produced 1,205. All 4,901 were ordinary clean borrowed inputs; their average encoded lengths were about 113.9 and 121.2 bytes respectively. No leading-parent, dirty, current-directory, or invalid-encoding normalization input appeared. A disposable external-module fixture forced the path-like external branch and produced one canonical leading-parent hit among seven calls, but that branch represented about 0.02% of the combined sample and the conditional count was only one; it is evidence that the shape exists, not a stable frequency estimate.
@@ -52,6 +58,8 @@ The final descendant strict-`String` variants were effectively equal in the full
 
 The final Windows implementation adds a native-separator-aware root and common-prefix scan over borrowed canonical inputs, while dirty and mixed-separator inputs retain the normalization fallback. Native Windows-MSVC GitHub Actions generation run [`29181673809`](https://github.com/hyf0/sugar_path/actions/runs/29181673809) produced the final-API Windows evidence; ordinary PR CI checks the committed MSVC snapshot via `cargo allocs`: canonical native descendant `relative -> Cow<Path>` is zero allocations, `Cow::into_owned` is one, descendant and upward ordinary final-`String` composition are one, clean `PathBuf::into_slash` is zero, and clean owned cwd reuse has no fresh allocation plus one growth reallocation. Ordinary and verbatim same-share and different-share UNC rows each construct only their one required owned result. The scenario matrix inside the allocation runner still covers drive, UNC, verbatim, device, mixed-separator, invalid-wide, and root-relative cancellation controls; exact tests cover the additional components that cannot be represented as standalone relative paths. This goal did not authorize or run Docker, Wine, or containers; non-container GNU/MSVC cross-checks and native Actions supply the final build and allocation evidence.
 
+Internally produced slash-form relative outcomes are valid UTF-8. On Windows, their owned `String` can therefore replace only ASCII `/` bytes with `\` in its existing allocation before becoming a `PathBuf`; a pointer-and-capacity test locks both exact Unicode output and buffer reuse without unsafe code. Borrowed slash conversion retains its single-pass replacement path because a `memchr` precheck followed by `str::replace` regressed the existing owned-result benchmark by 12.68% in Linux CodSpeed. Invalid native encoding still follows the strict failure or explicit lossy paths before this helper. GNU/MSVC cross-compilation proves target coverage only, so native Windows correctness, allocation snapshots, and timing remain the acceptance gate for the owned-buffer reuse.
+
 ## Benchmark audit findings
 
 The pre-baseline suite aggregated many paths into unnamed batch costs, did not declare throughput, black-boxed outputs but not inputs, and used Unix-shaped absolute fixtures on Windows. `relative_with_dots` normalized both inputs before calling `relative`, so it never exercised the named dot slow path. `as_path_vs_path_new` measured both operations in one closure and could not compare them. These benchmarks could detect very large changes but could not attribute smaller allocation or branch regressions.
@@ -194,7 +202,7 @@ The pinned [`rust-performance-improvement-plan` at `fe59e3d`](https://github.com
 | Reserve capacity, reuse collections, and avoid `clone`/`to_owned` | Adopted where ownership proves reuse. The final `into_normalized` and strict `into_slash` operations reuse clean owned buffers, and an owned cwd moves into `absolutize_with`; dirty normalization still builds the exact transformed representation rather than forcing mutation into the old buffer. |
 | Use `Cow`, `SmallVec`, and compact storage for short values | Retained where justified. `Cow<Path>` exposes allocation-free clean paths; normalization uses eight inline references, while lexical relative calculation uses 16 because the pinned Rolldown distribution is 10 components at p99 and 13 at maximum. Dedicated dirty and 24-level tests execute both spill paths. |
 | Avoid repeated `Path::to_string_lossy` | Adopted for valid-Unicode owned input through strict `into_slash` and for the ordinary relative-to-final-`String` composition. The earlier public fused method was removed; invalid native encoding still requires explicit lossy re-encoding, with byte- and wide-unit tests proving the fallback semantics. |
-| Add `#[inline]`, `#[cold]`, unchecked indexing, or compiler assertions | Use only with generated-code or profile evidence. Tiny classifiers are inline; no current profile justifies unsafe indexing or a permanent cold-path layout, and path lengths are too small to assume bounds checks dominate. |
+| Add `#[inline]`, `#[cold]`, unchecked indexing, or compiler assertions | Use only with generated-code or profile evidence. The normalization classifier is outlined because generated code and CodSpeed favor that separation, while the rebuilding core and leading-parent helper remain eligible for inlining. No current profile justifies unsafe indexing or marking the general rebuilding path cold. |
 | De-monomorphize generic functions | `relative(base: impl AsRef<Path>)` is a candidate for compile-time and binary-size measurement, not an assumed runtime optimization. Keep the generic shim until `cargo llvm-lines` or binary-size evidence shows meaningful duplication. |
 | Faster hashers, sorting changes, regex caching, and parallelism | Not applicable to the current algorithms: SugarPath has no hash table, sorting, regex, or independent work large enough to repay coordination overhead. |
 | Build flags such as one codegen unit, LTO, mimalloc, and `target-cpu=native` | The benchmark profile uses one codegen unit, fat LTO, and Rolldown's pinned mimalloc. Shared baselines deliberately reject `target-cpu=native`; target-specific SIMD experiments require a separate benchmark identity. |

benchmarks/allocations/x86_64-pc-windows-msvc.snap

@@ -16,7 +16,7 @@ Each row measures one operation after allocation-capable setup and an untracked
 | normalize / collapsing input -> normalized path | 1 | 0 | 0 |
 | normalize / noncanonical leading parents -> normalized path | 1 | 0 | 0 |
 | normalize / canonical leading parents -> normalized path | 0 | 0 | 0 |
-| normalize / clean invalid encoding -> normalized path | 1 | 0 | 0 |
+| normalize / clean invalid encoding -> normalized path | 0 | 0 | 0 |
 | normalize / owned clean input via borrowed receiver -> PathBuf | 1 | 0 | 0 |
 | normalize / owned clean input via owned receiver -> PathBuf | 0 | 0 | 0 |
 | normalize / owned dirty input via borrowed receiver -> PathBuf | 1 | 0 | 0 |
@@ -26,7 +26,7 @@ Each row measures one operation after allocation-capable setup and an untracked
 | normalize / owned collapsing input via borrowed receiver -> PathBuf | 2 | 0 | 0 |
 | normalize / owned collapsing input via owned receiver -> PathBuf | 1 | 0 | 0 |
 | normalize / owned invalid input via borrowed receiver -> PathBuf | 1 | 0 | 0 |
-| normalize / owned invalid input via owned receiver -> PathBuf | 1 | 0 | 0 |
+| normalize / owned invalid input via owned receiver -> PathBuf | 0 | 0 | 0 |
 | pipeline / dirty join via borrowed receiver -> PathBuf | 2 | 0 | 2 |
 | pipeline / dirty join via owned receiver -> PathBuf | 2 | 0 | 2 |
 | pipeline / dirty join via borrowed receiver -> String | 3 | 0 | 2 |
@@ -52,13 +52,13 @@ Each row measures one operation after allocation-capable setup and an untracked
 | relative / p99-depth equal leading-parent inputs -> relative path | 1 | 0 | 0 |
 | 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 |
+| relative / dotted slow path -> relative path | 3 | 0 | 0 |
+| relative / invalid-encoding absolute target -> relative path (setup excluded) | 1 | 0 | 0 |
 | relative / invalid-encoding relative target -> relative path (setup excluded) | 1 | 0 | 0 |
-| Windows / absolute forward-slash inputs -> relative path | 2 | 0 | 0 |
+| Windows / absolute forward-slash inputs -> relative path | 1 | 0 | 0 |
 | Windows / same drive-relative prefix -> relative path | 1 | 0 | 0 |
 | Windows / different drive-relative prefix -> relative path | 4 | 0 | 0 |
-| Windows / root-relative inputs -> relative path | 3 | 0 | 0 |
+| Windows / root-relative inputs -> relative path | 1 | 0 | 0 |
 | Windows / ordinary UNC same share -> relative path | 1 | 0 | 0 |
 | Windows / ordinary UNC different share -> relative path | 1 | 0 | 0 |
 | Windows / verbatim UNC same share -> relative path | 1 | 0 | 0 |
@@ -101,7 +101,7 @@ Each row measures one operation after allocation-capable setup and an untracked
 | normalize / collapsing input -> normalized path | 6 | 0 |
 | normalize / noncanonical leading parents -> normalized path | 51 | 0 |
 | normalize / canonical leading parents -> normalized path | 0 | 0 |
-| normalize / clean invalid encoding -> normalized path | 70 | 0 |
+| normalize / clean invalid encoding -> normalized path | 0 | 0 |
 | normalize / owned clean input via borrowed receiver -> PathBuf | 58 | 0 |
 | normalize / owned clean input via owned receiver -> PathBuf | 0 | 0 |
 | normalize / owned dirty input via borrowed receiver -> PathBuf | 58 | 0 |
@@ -111,7 +111,7 @@ Each row measures one operation after allocation-capable setup and an untracked
 | normalize / owned collapsing input via borrowed receiver -> PathBuf | 7 | 0 |
 | normalize / owned collapsing input via owned receiver -> PathBuf | 6 | 0 |
 | normalize / owned invalid input via borrowed receiver -> PathBuf | 70 | 0 |
-| normalize / owned invalid input via owned receiver -> PathBuf | 70 | 0 |
+| normalize / owned invalid input via owned receiver -> PathBuf | 0 | 0 |
 | pipeline / dirty join via borrowed receiver -> PathBuf | 100 | 126 |
 | pipeline / dirty join via owned receiver -> PathBuf | 100 | 126 |
 | pipeline / dirty join via borrowed receiver -> String | 170 | 126 |
@@ -133,17 +133,17 @@ Each row measures one operation after allocation-capable setup and an untracked
 | relative / relative inputs -> relative path | 31 | 0 |
 | relative / dotted relative inputs -> relative path | 23 | 0 |
 | relative / equal leading-parent inputs -> relative path | 23 | 0 |
-| relative / unequal leading-parent inputs -> relative path | 383 | 276 |
+| relative / unequal leading-parent inputs -> relative path | 60 | 0 |
 | relative / p99-depth equal leading-parent inputs -> relative path | 18 | 0 |
-| relative / p99-depth unequal leading-parent inputs -> relative path | 427 | 276 |
+| relative / p99-depth unequal leading-parent inputs -> relative path | 87 | 0 |
 | relative / current-directory spellings -> relative path | 0 | 0 |
-| relative / dotted slow path -> relative path | 195 | 0 |
-| relative / invalid-encoding absolute target -> relative path (setup excluded) | 139 | 0 |
+| relative / dotted slow path -> relative path | 164 | 0 |
+| relative / invalid-encoding absolute target -> relative path (setup excluded) | 48 | 0 |
 | relative / invalid-encoding relative target -> relative path (setup excluded) | 24 | 0 |
-| Windows / absolute forward-slash inputs -> relative path | 56 | 0 |
+| Windows / absolute forward-slash inputs -> relative path | 28 | 0 |
 | Windows / same drive-relative prefix -> relative path | 18 | 0 |
 | Windows / different drive-relative prefix -> relative path | 178 | 0 |
-| Windows / root-relative inputs -> relative path | 51 | 0 |
+| Windows / root-relative inputs -> relative path | 18 | 0 |
 | Windows / ordinary UNC same share -> relative path | 27 | 0 |
 | Windows / ordinary UNC different share -> relative path | 36 | 0 |
 | Windows / verbatim UNC same share -> relative path | 27 | 0 |

benchmarks/allocations/x86_64-unknown-linux-gnu.snap

@@ -16,7 +16,7 @@ Each row measures one operation after allocation-capable setup and an untracked
 | normalize / collapsing input -> normalized path | 1 | 0 | 0 |
 | normalize / noncanonical leading parents -> normalized path | 1 | 0 | 0 |
 | normalize / canonical leading parents -> normalized path | 0 | 0 | 0 |
-| normalize / clean invalid encoding -> normalized path | 1 | 0 | 0 |
+| normalize / clean invalid encoding -> normalized path | 0 | 0 | 0 |
 | normalize / owned clean input via borrowed receiver -> PathBuf | 1 | 0 | 0 |
 | normalize / owned clean input via owned receiver -> PathBuf | 0 | 0 | 0 |
 | normalize / owned dirty input via borrowed receiver -> PathBuf | 1 | 0 | 0 |
@@ -26,7 +26,7 @@ Each row measures one operation after allocation-capable setup and an untracked
 | normalize / owned collapsing input via borrowed receiver -> PathBuf | 2 | 0 | 0 |
 | normalize / owned collapsing input via owned receiver -> PathBuf | 1 | 0 | 0 |
 | normalize / owned invalid input via borrowed receiver -> PathBuf | 1 | 0 | 0 |
-| normalize / owned invalid input via owned receiver -> PathBuf | 1 | 0 | 0 |
+| normalize / owned invalid input via owned receiver -> PathBuf | 0 | 0 | 0 |
 | pipeline / dirty join via borrowed receiver -> PathBuf | 2 | 0 | 2 |
 | pipeline / dirty join via owned receiver -> PathBuf | 2 | 0 | 2 |
 | pipeline / dirty join via borrowed receiver -> String | 3 | 0 | 2 |
@@ -50,7 +50,7 @@ Each row measures one operation after allocation-capable setup and an untracked
 | 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 |
+| relative / invalid-encoding absolute target -> relative path (setup excluded) | 1 | 0 | 3 |
 | relative / invalid-encoding relative target -> relative path (setup excluded) | 1 | 0 | 0 |
 | to_slash / native path -> slash text | 0 | 0 | 0 |
 | to_slash_lossy / invalid encoding -> lossy slash text (setup excluded) | 1 | 0 | 1 |
@@ -88,7 +88,7 @@ Each row measures one operation after allocation-capable setup and an untracked
 | normalize / collapsing input -> normalized path | 6 | 0 |
 | normalize / noncanonical leading parents -> normalized path | 51 | 0 |
 | normalize / canonical leading parents -> normalized path | 0 | 0 |
-| normalize / clean invalid encoding -> normalized path | 66 | 0 |
+| normalize / clean invalid encoding -> normalized path | 0 | 0 |
 | normalize / owned clean input via borrowed receiver -> PathBuf | 56 | 0 |
 | normalize / owned clean input via owned receiver -> PathBuf | 0 | 0 |
 | normalize / owned dirty input via borrowed receiver -> PathBuf | 58 | 0 |
@@ -98,7 +98,7 @@ Each row measures one operation after allocation-capable setup and an untracked
 | normalize / owned collapsing input via borrowed receiver -> PathBuf | 7 | 0 |
 | normalize / owned collapsing input via owned receiver -> PathBuf | 6 | 0 |
 | normalize / owned invalid input via borrowed receiver -> PathBuf | 66 | 0 |
-| normalize / owned invalid input via owned receiver -> PathBuf | 66 | 0 |
+| normalize / owned invalid input via owned receiver -> PathBuf | 0 | 0 |
 | pipeline / dirty join via borrowed receiver -> PathBuf | 96 | 115 |
 | pipeline / dirty join via owned receiver -> PathBuf | 96 | 115 |
 | pipeline / dirty join via borrowed receiver -> String | 164 | 115 |
@@ -117,12 +117,12 @@ Each row measures one operation after allocation-capable setup and an untracked
 | relative / relative inputs -> relative path | 31 | 0 |
 | relative / dotted relative inputs -> relative path | 23 | 0 |
 | relative / equal leading-parent inputs -> relative path | 23 | 0 |
-| relative / unequal leading-parent inputs -> relative path | 203 | 250 |
+| relative / unequal leading-parent inputs -> relative path | 60 | 0 |
 | relative / p99-depth equal leading-parent inputs -> relative path | 18 | 0 |
-| relative / p99-depth unequal leading-parent inputs -> relative path | 220 | 388 |
+| relative / p99-depth unequal leading-parent inputs -> relative path | 87 | 0 |
 | relative / current-directory spellings -> relative path | 0 | 0 |
 | relative / dotted slow path -> relative path | 31 | 0 |
-| relative / invalid-encoding absolute target -> relative path (setup excluded) | 93 | 112 |
+| relative / invalid-encoding absolute target -> relative path (setup excluded) | 8 | 112 |
 | relative / invalid-encoding relative target -> relative path (setup excluded) | 22 | 0 |
 | to_slash / native path -> slash text | 0 | 0 |
 | to_slash_lossy / invalid encoding -> lossy slash text (setup excluded) | 66 | 132 |