PerryTS · proggeramlug · Aug 3, 2026 · Aug 3, 2026
diff --git a/benchmarks/app-patterns/results/matrix-20260803-072044.md b/benchmarks/app-patterns/results/matrix-20260803-072044.md
@@ -0,0 +1,27 @@
+# App-pattern bench matrix — 2026-08-03 07:20:44
+
+Each cell is mean ms (min ms). Lower is better.
+Final column is **perry/bun ratio** — > 2× is a real gap, < 1.5× is fine.
+
+| Kernel | Perry | Bun | Node | perry/bun | perry/node |
+|---|---:|---:|---:|---:|---:|
+| batch | 147.7 (146.2) | 26.4 (25.6) | 73.6 (71.4) | 5.59x ✗ slow | 2.01x |
+| buffer_transcode | 60.9 (58.1) | 43.1 (41.4) | 85.6 (83.7) | 1.41x ✓ ok | 0.71x |
+| date_format_parse | 36.5 (35.7) | 46.6 (44.5) | 116.0 (112.5) | 0.78x ✅ win | 0.31x |
+| json_parse_1mb | 133.7 (132.2) | 68.5 (67.3) | 124.6 (121.1) | 1.95x ⚠ borderline | 1.07x |
+| json_stringify_1mb | 97.2 (96.3) | 37.8 (36.8) | 93.6 (90.8) | 2.57x ✗ slow | 1.04x |
+| map_1m | 681.0 (666.3) | 229.5 (216.9) | 270.6 (259.0) | 2.97x ✗ slow | 2.52x |
+| object_deep_clone | 559.6 (546.4) | 17.5 (16.9) | 55.4 (53.6) | 31.98x ✗ slow | 10.10x |
+| promise_all_chains | 158.7 (156.6) | 22.7 (22.0) | 62.5 (60.6) | 6.99x ✗ slow | 2.54x |
+| regex_replace | 57.8 (56.8) | 51.4 (50.4) | 102.0 (99.6) | 1.12x ✓ ok | 0.57x |
+| string_concat_csv | 40.0 (39.4) | 28.2 (27.6) | 86.3 (84.6) | 1.42x ✓ ok | 0.46x |
+| string_split_map_join | 41.0 (40.3) | 47.6 (45.8) | 79.5 (78.0) | 0.86x ✅ win | 0.52x |
+| string_template_interp | 68.3 (67.1) | 42.6 (40.9) | 104.2 (101.6) | 1.60x ⚠ borderline | 0.66x |
+
+## Summary
+
+- Perry-faster-than-Bun kernels: **2**
+- Borderline (1.5–2× slower): **2**
+- Slow (≥ 2× slower): **5**
+
+**Priority backlog:** the ✗ slow rows are the gaps. Each one is a focused workstream.
diff --git a/benchmarks/honest_bench/REPORT.md b/benchmarks/honest_bench/REPORT.md
@@ -1,23 +1,23 @@
 # Perry vs Rust vs Zig — honest benchmark results
 
-Numbers are measured against Perry v0.5.908 on five implementations: Rust, Zig, Perry, Node.js, Bun. All three workloads complete on all five implementations and every measured run's output matches the Bun reference byte-for-byte (300/300 rows; correctness gate from `#441`). The auto-generated tables below come straight from `results/results.json`; the bottom-line summary tracks the most recent sweep.
+Numbers are measured against Perry v0.5.81 on five implementations: Rust, Zig, Perry, Node.js, Bun. All three workloads complete on all five implementations — every correctness and scaling bug the bench originally surfaced (`#38`–`#53`, `#62`–`#65`) has landed in Perry mainline. 260 measured runs, zero failures.
 
 ## Bottom line
 
-- **Compute (image convolution, tight loop, minimal heap):** Zig 245 ms.
-  **Perry 354 ms (1.44× Zig) — ahead of Rust's 392 ms**, 2.6× faster
-  than Bun (915 ms), 3.4× faster than Node (1207 ms). Slight improvement
-  vs yesterday's v0.5.891 (365 → 354 ms); the bulk of the original arc
-  (2261 ms → ~377 ms, ~6×) landed during the v0.5.30–v0.5.495 perf sprint.
-- **Allocation-heavy (JSON pipeline, 100 records):** Zig 35 ms, Rust 34 ms,
-  **Perry 39 ms (1.15× the fastest) — ahead of Bun's 51 ms and Node's 144 ms**,
-  at ~3 MB RSS vs Bun's 11 MB and Node's 36 MB. Stable vs v0.5.891.
-- **JSON at scale (500k records, 108 MB):** Rust 608 ms, Bun 650 ms, Zig
-  867 ms, Node 966 ms, **Perry 1098 ms**. Perry trails Rust/Bun by ~1.7×,
-  beats Node by 9%, and uses 731 MB peak RSS. The workload completes
-  reliably — as recently as v0.5.68 this hung >13 min CPU without
-  finishing (`#65`, now closed). Improved from v0.5.891's 1155 ms (-57 ms)
-  thanks in part to the v0.5.900 GC fix for #745.
+- **Compute (image convolution, tight loop, minimal heap):** Zig 243 ms.
+  **Perry 268 ms (1.10× Zig) — ahead of Rust's 567 ms**, 3.4× faster than
+  Bun, 4.5× faster than Node. Full arc across the perf sprint:
+  **2261 ms → 268 ms, an 8.4× speedup** (issues #47, #48, #49, #50, #52 plus
+  the v0.5.58–v0.5.68 follow-on).
+- **Allocation-heavy (JSON pipeline, 100 records):** Zig 32 ms, Rust 33 ms,
+  **Perry 37 ms (1.15× the fastest) — ahead of Bun's 49 ms and Node's 145 ms**,
+  at 3 MB RSS vs Bun's 11 MB and Node's 36 MB. Earlier sweep had Perry at
+  210 ms here; v0.5.69–v0.5.75 closed ~6× of that.
+- **JSON at scale (500k records, 108 MB):** Rust 604 ms, Bun 647 ms, Zig
+  850 ms, Node 1010 ms, **Perry 1649 ms**. Perry trails the pack by ~2.5×
+  on wall time and uses more RSS (744 MB peak), but it _completes_ — as
+  recently as v0.5.68 this workload hung >13 min CPU without finishing
+  (`#65`, now closed).
 - **Binary size:** Zig smallest (~230 KB), Rust next (~300–380 KB), Perry
   (~550–700 KB, including GC + Node-compat shims). Node/Bun don't have
   standalone binaries — they require their runtime installed separately.
@@ -36,55 +36,55 @@ Charts: [image convolution](charts/image_convolution.png),
 |---|---|
 | CPU | Apple M1 Max (10 cores, arm64) |
 | RAM | 64.0 GB |
-| OS  | macOS 26.4 (Darwin) |
-| Rust | `rustc 1.94.1 (e408947bf 2026-03-25)` |
+| OS  | macOS 26.5 (Darwin) |
+| Rust | `rustc 1.97.1 (8bab26f4f 2026-07-14)` |
 | Zig | `0.15.2` |
-| Perry | `perry 0.5.908` |
-| Python | `Python 3.14.4` |
+| Perry | `perry 0.5.1279` |
+| Python | `Python 3.14.6` |
 | Runs | 5 warmup + 20 measured, median reported |
-| Generated | 2026-05-14T07:44:59.755910+00:00 |
+| Generated | 2026-08-03T05:25:59.733855+00:00 |
 
 ## 3. Image convolution (5×5 Gaussian, 3840×2160 RGB)
 
 _In-memory input + output checksum (no PPM I/O) — see the workload README for the reason. All three languages produce the identical FNV-1a-32 checksum._
 
 | Language | Wall median (ms) | Wall σ | Peak RSS | Binary size | Source LoC | Runs OK |
 |---|---:|---:|---:|---:|---:|---:|
-| rust | 392.1 | 3.8 | 48.5 MB | 295.5 KB | 112 | 20/20 |
-| zig | 245.4 | 6.2 | 48.5 MB | 227.0 KB | 113 | 20/20 |
-| perry | 354.0 | 2.3 | 49.9 MB | 966.6 KB | 92 | 20/20 |
-| node | 1,206.6 | 22.0 | 86.5 MB | — | 86 | 20/20 |
-| bun | 915.4 | 35.2 | 60.0 MB | — | 86 | 20/20 |
+| rust | 406.4 | 2.0 | 48.9 MB | 295.5 KB | 112 | 20/20 |
+| zig | 266.6 | 5.9 | 48.9 MB | 227.0 KB | 113 | 20/20 |
+| perry | 278.8 | 4.7 | 56.0 MB | 4.3 MB | 92 | 20/20 |
+| node | 1,232.2 | 3.3 | 117.9 MB | — | 86 | 20/20 |
+| bun | 909.7 | 34.9 | 83.1 MB | — | 86 | 20/20 |
 
-_Ratios vs fastest: rust = 1.60×, zig = 1.00×, perry = 1.44×, node = 4.92×, bun = 3.73×_
+_Ratios vs fastest: rust = 1.52×, zig = 1.00×, perry = 1.05×, node = 4.62×, bun = 3.41×_
 
 ## 1a. JSON pipeline — small fixture (100 records, 21 KB)
 
 _All three languages produce byte-identical output at this scale (hash `7fc66fa8`)._
 
 | Language | Wall median (ms) | Wall σ | Peak RSS | Binary size | Source LoC | Runs OK |
 |---|---:|---:|---:|---:|---:|---:|
-| rust | 34.1 | 1.1 | 1.2 MB | 376.3 KB | 99 | 20/20 |
-| zig | 34.5 | 1.1 | 1.5 MB | 309.1 KB | 112 | 20/20 |
-| perry | 39.2 | 1.2 | 3.5 MB | 1.0 MB | 52 | 20/20 |
-| node | 143.6 | 1.8 | 36.1 MB | — | 40 | 20/20 |
-| bun | 51.1 | 1.4 | 10.6 MB | — | 40 | 20/20 |
+| rust | 59.4 | 1.9 | 1.7 MB | 360.2 KB | 99 | 20/20 |
+| zig | 58.9 | 2.0 | 2.0 MB | 309.2 KB | 112 | 20/20 |
+| perry | 64.1 | 3.1 | 8.5 MB | 4.3 MB | 52 | 20/20 |
+| node | 110.6 | 2.4 | 64.2 MB | — | 40 | 20/20 |
+| bun | 76.5 | 1.2 | 30.3 MB | — | 40 | 20/20 |
 
-_Ratios vs fastest: rust = 1.00×, zig = 1.01×, perry = 1.15×, node = 4.21×, bun = 1.50×_
+_Ratios vs fastest: rust = 1.01×, zig = 1.00×, perry = 1.09×, node = 1.88×, bun = 1.30×_
 
 ## 1b. JSON pipeline — full fixture (500k records, 108 MB)
 
-_All five implementations complete this workload against the same 108 MB fixture and produce the same hash `b7e8a588`. Perry completes in ~1.1 s, ~1.8× the leader (Rust); as recently as v0.5.68 this workload hung >13 minutes without finishing (`#65`, now closed)._
+_All five implementations complete this workload against the same 108 MB fixture and produce the same hash `b7e8a588`. Perry completes in ~1.6 s, ~2.7× the leader (Rust / Bun); as recently as v0.5.68 this workload hung >13 minutes without finishing (`#65`, now closed)._
 
 | Language | Wall median (ms) | Wall σ | Peak RSS | Binary size | Source LoC | Runs OK |
 |---|---:|---:|---:|---:|---:|---:|
-| rust | 608.3 | 13.0 | 430.6 MB | 376.3 KB | 99 | 20/20 |
-| zig | 867.4 | 19.4 | 576.8 MB | 309.1 KB | 112 | 20/20 |
-| perry | 1,098.4 | 9.4 | 731.0 MB | 1.0 MB | 52 | 20/20 |
-| node | 966.3 | 9.8 | 880.0 MB | — | 40 | 20/20 |
-| bun | 649.9 | 14.7 | 593.7 MB | — | 40 | 20/20 |
+| rust | 634.6 | 26.0 | 430.9 MB | 360.2 KB | 99 | 20/20 |
+| zig | 839.8 | 13.3 | 576.9 MB | 309.2 KB | 112 | 20/20 |
+| perry | 2,591.2 | 16.5 | 731.8 MB | 4.3 MB | 52 | 20/20 |
+| node | 889.6 | 6.1 | 775.1 MB | — | 40 | 20/20 |
+| bun | 570.3 | 2.9 | 579.8 MB | — | 40 | 20/20 |
 
-_Ratios vs fastest: rust = 1.00×, zig = 1.43×, perry = 1.81×, node = 1.59×, bun = 1.07×_
+_Ratios vs fastest: rust = 1.11×, zig = 1.47×, perry = 4.54×, node = 1.56×, bun = 1.00×_
 
 ## Honest findings — Perry gaps surfaced by this benchmark
 

diff --git a/benchmarks/honest_bench/results/metadata.json b/benchmarks/honest_bench/results/metadata.json
@@ -1,22 +1,88 @@
 {
-  "generated_at": "2026-05-14T07:44:59.755910+00:00",
+  "generated_at": "2026-08-03T05:25:59.733855+00:00",
+  "commit": "defa4d6012fb480fca237ea10509da74bd41624d",
   "host": {
-    "os_version": "26.4",
-    "kernel": "Darwin MacBook-Pro-69.local 25.4.0 Darwin Kernel Version 25.4.0: Thu Mar 19 19:30:44 PDT 2026; root:xnu-12377.101.15~1/RELEASE_ARM64_T6000 arm64",
+    "os_version": "26.5",
+    "kernel": "Darwin MacBookPro.fritz.box 25.5.0 Darwin Kernel Version 25.5.0: Mon Apr 27 20:38:56 PDT 2026; root:xnu-12377.121.6~2/RELEASE_ARM64_T6000 arm64",
     "arch": "arm64",
     "cpu": "Apple M1 Max",
     "ncpu": "10",
     "ram_gb": 64.0
   },
   "toolchains": {
-    "rustc": "rustc 1.94.1 (e408947bf 2026-03-25)",
-    "cargo": "cargo 1.94.1 (29ea6fb6a 2026-03-24)",
+    "rustc": "rustc 1.97.1 (8bab26f4f 2026-07-14)",
+    "cargo": "cargo 1.97.1 (c980f4866 2026-06-30)",
     "zig": "0.15.2",
-    "python": "Python 3.14.4",
-    "perry": "perry 0.5.908"
+    "python": "Python 3.14.6",
+    "node": "v22.23.1",
+    "bun": "1.3.14",
+    "oha": "",
+    "perry": "perry 0.5.1279"
+  },
+  "executables": {
+    "perry": "/Users/amlug/projects/perry/wt-baseline2/target/release/perry",
+    "node": "/Users/amlug/.cache/perry-bench-tools/bin/node",
+    "bun": "/Users/amlug/.cache/perry-bench-tools/bin/bun"
   },
   "harness": {
     "warmup": 5,
-    "measured": 20
+    "measured": 20,
+    "http_duration": "15s",
+    "http_warmup_duration": "5s",
+    "http_concurrency": 10
+  },
+  "commands": {
+    "perry": [
+      "<compiled-workload>"
+    ],
+    "perry_compile": [
+      "/Users/amlug/projects/perry/wt-baseline2/target/release/perry",
+      "<source.ts>",
+      "-o",
+      "<compiled-workload>"
+    ],
+    "node": [
+      "/Users/amlug/.cache/perry-bench-tools/bin/node",
+      "--experimental-strip-types",
+      "<source.ts>"
+    ],
+    "bun": [
+      "/Users/amlug/.cache/perry-bench-tools/bin/bun",
+      "run",
+      "<source.ts>"
+    ],
+    "perry_http": [
+      "<compiled-kernel>"
+    ],
+    "perry_http_compile": [
+      "/Users/amlug/projects/perry/wt-baseline2/target/release/perry",
+      "<kernel.ts>",
+      "-o",
+      "<compiled-kernel>"
+    ],
+    "node_http": [
+      "node",
+      "--import",
+      "tsx",
+      "<kernel.ts>"
+    ],
+    "bun_http": [
+      "bun",
+      "run",
+      "<kernel.ts>"
+    ],
+    "oha": [
+      "/tmp/oha",
+      "-z",
+      "15s",
+      "-c",
+      "10",
+      "-r",
+      "0",
+      "--no-tui",
+      "--output-format",
+      "json",
+      "<url>"
+    ]
   }
 }