[Autoloop: tsb-perf-evolve] Iteration 41: inverse IEEE-754 gather — avoid random vals[] access

Replace the gather loop's `vals[origIdx]` random access into the JS array
with an inverse IEEE-754 transform that reconstructs the float value directly
from the already-sorted AoS key words (keyLo, keyHi) in srcBuf. The sorted
AoS buffer is hot in cache after the final scatter pass; the original vals[]
array is cold after the init loop. Replacing the random read with sequential
reads + ~6 arithmetic ops should reduce cache pressure in the gather phase.

Run: https://github.com/githubnext/tsessebe/actions/runs/25712799061

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: github-actions[bot]

src/core/series.ts

@@ -901,16 +901,34 @@ export class Series<T extends Scalar = Scalar> {
       if (allNumeric) {
         if (ascending) {
           for (let i = 0, si = 0; i < finCount; i++, si += 3) {
-            const idx = srcBuf[si]!;
-            perm[pos] = idx;
-            outData[pos] = vals[idx] as T;
+            const origIdx = srcBuf[si]!;
+            const keyLo = srcBuf[si + 1]!;
+            const keyHi = srcBuf[si + 2]!;
+            perm[pos] = origIdx;
+            if (keyHi & 0x80000000) {
+              _fvalsU32[0] = keyLo;
+              _fvalsU32[1] = (keyHi ^ 0x80000000) >>> 0;
+            } else {
+              _fvalsU32[0] = (~keyLo) >>> 0;
+              _fvalsU32[1] = (~keyHi) >>> 0;
+            }
+            outData[pos] = _fvals[0] as T;
             pos = pos + 1;
           }
         } else {
           for (let i = finCount - 1, si = (finCount - 1) * 3; i >= 0; i--, si -= 3) {
-            const idx = srcBuf[si]!;
-            perm[pos] = idx;
-            outData[pos] = vals[idx] as T;
+            const origIdx = srcBuf[si]!;
+            const keyLo = srcBuf[si + 1]!;
+            const keyHi = srcBuf[si + 2]!;
+            perm[pos] = origIdx;
+            if (keyHi & 0x80000000) {
+              _fvalsU32[0] = keyLo;
+              _fvalsU32[1] = (keyHi ^ 0x80000000) >>> 0;
+            } else {
+              _fvalsU32[0] = (~keyLo) >>> 0;
+              _fvalsU32[1] = (~keyHi) >>> 0;
+            }
+            outData[pos] = _fvals[0] as T;
             pos = pos + 1;
           }
         }
@@ -926,16 +944,36 @@ export class Series<T extends Scalar = Scalar> {
       if (allNumeric) {
         if (ascending) {
           for (let i = 0, si = 0; i < finCount; i++, si += 3) {
-            const idx = srcBuf[si]!;
-            perm[pos] = idx;
-            outData[pos] = vals[idx] as T;
+            const origIdx = srcBuf[si]!;
+            const keyLo = srcBuf[si + 1]!;
+            const keyHi = srcBuf[si + 2]!;
+            perm[pos] = origIdx;
+            // Reverse the IEEE-754 sign-transform to recover the original float bits,
+            // avoiding a random read into the JS values array.
+            if (keyHi & 0x80000000) {
+              _fvalsU32[0] = keyLo;
+              _fvalsU32[1] = (keyHi ^ 0x80000000) >>> 0;
+            } else {
+              _fvalsU32[0] = (~keyLo) >>> 0;
+              _fvalsU32[1] = (~keyHi) >>> 0;
+            }
+            outData[pos] = _fvals[0] as T;
             pos = pos + 1;
           }
         } else {
           for (let i = finCount - 1, si = (finCount - 1) * 3; i >= 0; i--, si -= 3) {
-            const idx = srcBuf[si]!;
-            perm[pos] = idx;
-            outData[pos] = vals[idx] as T;
+            const origIdx = srcBuf[si]!;
+            const keyLo = srcBuf[si + 1]!;
+            const keyHi = srcBuf[si + 2]!;
+            perm[pos] = origIdx;
+            if (keyHi & 0x80000000) {
+              _fvalsU32[0] = keyLo;
+              _fvalsU32[1] = (keyHi ^ 0x80000000) >>> 0;
+            } else {
+              _fvalsU32[0] = (~keyLo) >>> 0;
+              _fvalsU32[1] = (~keyHi) >>> 0;
+            }
+            outData[pos] = _fvals[0] as T;
             pos = pos + 1;
           }
         }