Retune default ladder scale to 2.5; fix inverted native speedup ratioin benchmark summary

.github/workflows/tuning.yml

@@ -1,11 +1,12 @@
 name: RD tuning sweep
 
 # Parallel parameter sweep for known-flaws A3/A4. Dispatch builds the codec
-# with 8 different ladder/divisor combinations simultaneously. Each job
+# with 9 different ladder/divisor combinations simultaneously. Each job
 # produces an RD table in the step summary. Compare all tables to pick
 # winning constants, then bake them into v2_core.cpp as the new defaults.
-# Stock values: scale=1.5, divisor=16.0 (retuned after the photo-pattern RD
-# sweep; divisor 16 adds intermediate lossy ladder steps with no regressions).
+# Stock values: scale=2.5, divisor=16.0 (retuned after the photo-pattern RD
+# sweeps; divisor 16 adds intermediate lossy ladder steps and scale 2.5 yields
+# the finest usable ladder - nine lossy steps on natural content).
 
 on:
   workflow_dispatch:
@@ -21,6 +22,9 @@ jobs:
       matrix:
         include:
           - label: stock
+            scale: "2.5"
+            divisor: "16.0"
+          - label: legacy-ladder
             scale: "1.5"
             divisor: "16.0"
           - label: steep-ladder
@@ -41,8 +45,8 @@ jobs:
           - label: gentle-divisor
             scale: "1.5"
             divisor: "4.0"
-          - label: steep-divisor
-            scale: "2.5"
+          - label: steepest
+            scale: "3.5"
             divisor: "16.0"
     steps:
       - uses: actions/checkout@v6

CHANGELOG.md

@@ -4,6 +4,14 @@ All notable WIMF changes are recorded here. The project follows semantic version
 
 ## 2.2.0 - Unreleased
 
+- Retuned the default quality ladder from scale 1.5 to 2.5 (divisor stays 16)
+  based on the photo/natural RD sweeps: the lossy ladder gains a ninth usable
+  step on natural content and every quality index produces smaller files at
+  smoothly increasing PSNR. Same-quality-index output is more compressed and
+  slightly softer than 2.1; lossless output is unchanged. Old files decode
+  identically - the container format is untouched.
+- Fixed the codec benchmark summary printing "Native speedup 0x": the ratio
+  was computed inverted and is now guarded against zero reference rates.
 - Upgraded the scalar CRC-32 to slice-by-8: all eight slice tables are derived
   from the polynomial at compile time, and the bulk loop consumes eight bytes
   per iteration instead of one (typically 4-6x table throughput on every

benchmarks/codec_summary.py

@@ -26,9 +26,11 @@ def main() -> int:
     )
     print(native_row)
     print(reference_row)
-    if native["encode_mpx_s"] > 0 and native["decode_mpx_s"] > 0:
-        encode_speedup = reference["encode_mpx_s"] / native["encode_mpx_s"]
-        decode_speedup = reference["decode_mpx_s"] / native["decode_mpx_s"]
+    if reference["encode_mpx_s"] > 0 and reference["decode_mpx_s"] > 0:
+        # Speedup is native relative to the Python reference; skip the row
+        # when a reference rate rounds to zero so the ratio stays finite.
+        encode_speedup = native["encode_mpx_s"] / reference["encode_mpx_s"]
+        decode_speedup = native["decode_mpx_s"] / reference["decode_mpx_s"]
         print(f"| Native speedup | {encode_speedup:.0f}x | {decode_speedup:.0f}x | - |")
     print()
     print("Each runner is different hardware: compare within a table, not across tables or runs.")

docs/known-flaws.md

@@ -33,9 +33,10 @@ determinism, corruption rejection, and memory guards all audited clean
   streams. Still missing: context modeling of residuals/subbands - the
   structural advantage modern image codecs exploit.
 - **[P1] A4 Coarse quantizer dead zone.** Largely fixed: monotonic scoring
-  (A4), the divisor retune 8.0→16.0, and the 0.9x quantizer sub-step in tile
-  scoring removed the non-monotonicity and the one-ladder-notch gap that showed
-  as a 34-43 dB jump on the photo pattern. Remaining coarseness is minor; the
+  (A4), the divisor retune 8.0→16.0, the 0.9x quantizer sub-step in tile
+  scoring, and the ladder scale retune 1.5→2.5 leave natural-content ladders
+  with nine smooth lossy steps. The photo pattern retains one Q5→Q6 jump from
+  per-tile energy bimodality (sharp edges next to flat regions); the
   structural win left is the context-modeled entropy stage (A3).
 - **[P2] A5 Per-tile framing overhead.** Default 128 px tiles give ~2.8k
   independent Zstd frames per 45 MP encode (per scored mode), with no

src/v2_core.cpp

@@ -18,12 +18,14 @@
 #include "zstd.h"
 #include "v2_simd.hpp"
 
-// Tunable codec constants. The ladder scale reproduces the historical quality
-// ladder; the scoring divisor default was retuned from 8.0 to 16.0 after an
-// RD sweep showed it adds intermediate lossy ladder steps with no regressions.
-// The tuning workflow overrides these via -D flags to sweep candidate curves.
+// Tunable codec constants. The scoring divisor and ladder scale were retuned
+// from the historical 8.0/1.5 after photo-pattern RD sweeps: divisor 16 adds
+// intermediate lossy ladder steps, and scale 2.5 yields the finest usable
+// ladder (nine lossy steps on natural content, smaller files at every quality
+// index). The tuning workflow overrides these via -D flags to sweep candidate
+// curves.
 #ifndef WIMF_LADDER_SCALE
-#define WIMF_LADDER_SCALE 1.5f
+#define WIMF_LADDER_SCALE 2.5f
 #endif
 #ifndef WIMF_SCORING_DIVISOR
 #define WIMF_SCORING_DIVISOR 16.0

tools/wimf_rd_sweep.cpp

@@ -22,7 +22,7 @@
 #include <vector>
 
 #ifndef WIMF_LADDER_SCALE
-#define WIMF_LADDER_SCALE 1.5f
+#define WIMF_LADDER_SCALE 2.5f
 #endif
 #ifndef WIMF_SCORING_DIVISOR
 #define WIMF_SCORING_DIVISOR 16.0

wiki/Known-Flaws.md

@@ -33,9 +33,10 @@ determinism, corruption rejection, and memory guards all audited clean
   streams. Still missing: context modeling of residuals/subbands - the
   structural advantage modern image codecs exploit.
 - **[P1] A4 Coarse quantizer dead zone.** Largely fixed: monotonic scoring
-  (A4), the divisor retune 8.0→16.0, and the 0.9x quantizer sub-step in tile
-  scoring removed the non-monotonicity and the one-ladder-notch gap that showed
-  as a 34-43 dB jump on the photo pattern. Remaining coarseness is minor; the
+  (A4), the divisor retune 8.0→16.0, the 0.9x quantizer sub-step in tile
+  scoring, and the ladder scale retune 1.5→2.5 leave natural-content ladders
+  with nine smooth lossy steps. The photo pattern retains one Q5→Q6 jump from
+  per-tile energy bimodality (sharp edges next to flat regions); the
   structural win left is the context-modeled entropy stage (A3).
 - **[P2] A5 Per-tile framing overhead.** Default 128 px tiles give ~2.8k
   independent Zstd frames per 45 MP encode (per scored mode), with no