Add parse-bench AST parity gate and fix Java > vs >> precpred

.github/workflows/parse-bench.yml

@@ -69,6 +69,15 @@ jobs:
           git -C /tmp/antlr-cleanroom/grammars-v4 fetch --depth 1 origin 284602b3f23ca54dc30778204ab7ae9e969145e9
           git -C /tmp/antlr-cleanroom/grammars-v4 checkout FETCH_HEAD
 
+      - name: Rust/Go AST parity (kotlin + java + trino)
+        run: |
+          python head/tools/parse-bench/run.py \
+            --languages kotlin,java,trino \
+            --runtimes rust-antlr,go-antlr \
+            --ast-check \
+            --quick \
+            --work-dir head/target/parse-bench-ast
+
       - name: Run base benchmark
         if: github.event_name == 'pull_request'
         run: |

README.md

@@ -450,6 +450,18 @@ python3 tools/parse-bench/run.py \
   --markdown target/parse-bench/results.md
 ```
 
+Add `--ast-check` to require byte-identical Rust/Go parse trees (no error nodes)
+before timing. Prefer that gate for fair comparisons; some C# Mono fixtures still
+diverge today (see `tools/parse-bench/README.md`). For a clean smoke:
+
+```bash
+python3 tools/parse-bench/run.py \
+  --languages kotlin,trino \
+  --runtimes rust-antlr,go-antlr \
+  --ast-check \
+  --quick
+```
+
 The report prints `min`/`avg` parse time and a ratio against `rust-antlr` for
 every fixture. Use `--quick` for a 3-iteration/1-warmup smoke run, or adjust
 `--iters`/`--warmups` for longer, lower-variance runs; add