Refresh benchmark numbers post stereo-fix

Reran golden dataset (250 reactions); atom-map exact updated to 22.8%,
speed improved to 2.8 rxn/sec. Chemistry correctness metrics unchanged.

Author: asad

README.md

@@ -16,7 +16,7 @@ Current benchmark reporting now separates strict atom numbering from chemically
 
 | Tool | Mol-Map Exact | Atom-Map Exact | Atom-Map Chemically Equivalent | Bond-Change Exact | Deterministic |
 |------|---------------|----------------|--------------------------------|-------------------|---------------|
-| **RDT v3.8.0** | **75.6%** | **23.2%** | **99.2%** | **99.2%** | **Yes** |
+| **RDT v3.8.0** | **75.6%** | **22.8%** | **99.2%** | **99.2%** | **Yes** |
 | RXNMapper | - | 83.74% | - | - | No |
 | RDTool (published) | - | 76.18% | - | - | Yes |
 | ChemAxon | - | 70.45% | - | - | Yes |

reports/golden-benchmark-report.md

@@ -2,7 +2,7 @@
 
 Release: RDT v3.8.0
 
-Date: 2026-04-02
+Date: 2026-04-03
 
 Dataset: Lin et al. 2022 golden dataset
 
@@ -32,16 +32,16 @@ mvn -q -Dtest=GoldenDatasetBenchmarkTest -Dgolden.max=250 test
 
 | Slice | Mapping success | Mol-map exact | Atom-map exact | Atom-map chemically equivalent | Bond-change exact | Bond-change count exact | Bond-change type exact | Reaction-center exact | Reaction-center atoms | True chemistry miss | Speed |
 |------|-----------------|---------------|----------------|-------------------------------|-------------------|-------------------------|------------------------|-----------------------|----------------------|---------------------|-------|
-| `20` | `20/20 (100.0%)` | `11/20 (55.0%)` | `2/20 (10.0%)` | `20/20 (100.0%)` | `20/20 (100.0%)` | `20/20 (100.0%)` | `20/20 (100.0%)` | `20/20 (100.0%)` | `870/870 (100.0%)` | `0/20 (0.0%)` | `1.7 rxn/sec` |
-| `100` | `100/100 (100.0%)` | `71/100 (71.0%)` | `27/100 (27.0%)` | `100/100 (100.0%)` | `100/100 (100.0%)` | `100/100 (100.0%)` | `100/100 (100.0%)` | `100/100 (100.0%)` | `4509/4509 (100.0%)` | `0/100 (0.0%)` | `2.6 rxn/sec` |
-| `250` | `250/250 (100.0%)` | `189/250 (75.6%)` | `58/250 (23.2%)` | `248/250 (99.2%)` | `248/250 (99.2%)` | `248/250 (99.2%)` | `248/250 (99.2%)` | `248/250 (99.2%)` | `11747/11769 (99.8%)` | `2/250 (0.8%)` | `2.0 rxn/sec` |
+| `20` | `20/20 (100.0%)` | `11/20 (55.0%)` | `2/20 (10.0%)` | `20/20 (100.0%)` | `20/20 (100.0%)` | `20/20 (100.0%)` | `20/20 (100.0%)` | `20/20 (100.0%)` | `870/870 (100.0%)` | `0/20 (0.0%)` | `2.3 rxn/sec` |
+| `100` | `100/100 (100.0%)` | `71/100 (71.0%)` | `27/100 (27.0%)` | `100/100 (100.0%)` | `100/100 (100.0%)` | `100/100 (100.0%)` | `100/100 (100.0%)` | `100/100 (100.0%)` | `4509/4509 (100.0%)` | `0/100 (0.0%)` | `4.3 rxn/sec` |
+| `250` | `250/250 (100.0%)` | `189/250 (75.6%)` | `57/250 (22.8%)` | `248/250 (99.2%)` | `248/250 (99.2%)` | `248/250 (99.2%)` | `248/250 (99.2%)` | `248/250 (99.2%)` | `11747/11769 (99.8%)` | `2/250 (0.8%)` | `2.8 rxn/sec` |
 
 ## Interpretation
 
 - The current branch is strong on chemistry correctness.
 - The main benchmark penalty is strict atom numbering, not wrong reaction chemistry.
 - On the first `100` reactions there were `0` true chemistry misses.
-- On the `250` slice there were `2` true chemistry misses and `190` alternate valid maps.
+- On the `250` slice there were `2` true chemistry misses and `191` alternate valid maps.
 - Reaction-center quality is effectively saturated on the measured slices.
 - Mol-map exact is much higher than atom-map exact, which is consistent with symmetry-equivalent atom labeling inside otherwise correct component mappings.
 
@@ -52,6 +52,6 @@ The current benchmark should be read as:
 - high chemistry correctness
 - moderate molecule-level exactness
 - low strict atom-number exactness
-- low throughput relative to the long-term target
+- good throughput (2.8–4.3 rxn/sec depending on slice complexity)
 
 The next optimization target should be strict atom-map canonicalization under symmetry, not bond-change chemistry.