feat: tweak perf flags

packages/cli/src/utils/node.ts

@@ -38,9 +38,16 @@ export function getNodeFlags(nodeMajorVersion: number): string[] {
     ];
   }
 
-  return [
+  const flags = [
     ...COMMON_NODE_FLAGS,
     // Disable optimizing compilers, keep the baseline compilers: sparkplug (JS), liftoff (WASM)
     '--max-opt=1',
   ];
+
+  if (nodeMajorVersion >= 22) {
+    // Disable concurrent marking for young generation (Minor Mark-Sweep); added in V8 12.x / Node 22
+    flags.push('--no-concurrent-minor-ms-marking');
+  }
+
+  return flags;
 }

test-apps/native/stability-harness.sh

@@ -0,0 +1,116 @@
+#!/usr/bin/env bash
+# =============================================================================
+# stability-harness.sh — Node.js flag stability assessment tool
+# =============================================================================
+#
+# PURPOSE
+#   Measures how stable (reproducible) performance measurements are under a
+#   given set of Node.js flags. Runs the perf test suite twice with identical
+#   flags and reports the coefficient of variation (CV = stdev/mean) for each
+#   measured scenario. Lower CV means more stable measurements.
+#
+# USAGE
+#   Run from test-apps/native or the repo root:
+#
+#     ./stability-harness.sh [node flags...]
+#
+#   Examples:
+#     ./stability-harness.sh                                        # current defaults only
+#     ./stability-harness.sh --expose-gc --max-opt=1               # specific flag set
+#     ./stability-harness.sh --expose-gc --max-opt=1 --some-flag   # testing a candidate
+#
+# HOW TO ASSESS A CANDIDATE FLAG
+#   The harness runs BOTH passes with the SAME flags — it is a stability probe,
+#   not an A/B test. To assess whether a flag helps:
+#
+#   1. Run the harness several times WITHOUT the flag (base configuration):
+#        ./stability-harness.sh --expose-gc --no-concurrent-sweeping --max-opt=1 --max-old-space-size=4096
+#      Record the "Weighted Average" values from both runs across all invocations.
+#
+#   2. Run the harness the same number of times WITH the candidate flag added:
+#        ./stability-harness.sh --expose-gc --no-concurrent-sweeping --max-opt=1 --max-old-space-size=4096 --some-flag
+#      Record the "Weighted Average" values.
+#
+#   3. Compare the two sets. A flag is worth adopting if it consistently lowers
+#      the average CV across multiple invocations. Single-run comparisons are
+#      unreliable — run at least 3 invocations per configuration.
+#
+#   4. NEVER run two harness invocations in parallel. Parallel runs share CPU
+#      resources and contaminate each other's measurements.
+#
+# OUTPUT INTERPRETATION
+#   "Stability / Weighted Average: X% => Y%"
+#     X% = mean-duration-weighted average CV of Run 1 (cold)
+#     Y% = mean-duration-weighted average CV of Run 2 (warm)
+#
+#   Both X and Y describe the SAME flag configuration. Run 2 is typically
+#   more stable because the JIT and allocator are warmer. Focus on Y% when
+#   comparing configurations; use X% to spot cold-start regressions.
+#
+#   Typical values with the current default flags (Node 22):
+#     Good:    < 1.0%
+#     Okay:    1.0% – 1.5%
+#     Noisy:   > 1.5%
+#
+# FLAGS ALREADY EVALUATED (as of 2026-06)
+#   Adopted (in packages/cli/src/utils/node.ts):
+#     --expose-gc                      always on; enables manual gc() between runs
+#     --no-concurrent-sweeping         all nodes; reduces GC thread interference
+#     --max-old-space-size=4096        all nodes; caps heap, reduces GC pressure variance
+#     --no-opt                         Node 18 only; disables all JIT optimisation
+#     --max-opt=1                      Node 19+; caps JIT at Sparkplug, blocks Turbofan/Maglev
+#     --no-concurrent-minor-ms-marking Node 22+; disables concurrent young-gen marking (V8 12.x+)
+#
+#   Ruled out (no reliable improvement):
+#     --no-concurrent-marking          negligible / noisy
+#     --no-incremental-marking         makes things worse
+#     --gc-interval=1000000            makes things worse
+#     --single-threaded-gc             neutral or slightly worse
+#     --initial-old-space-size=4096    dramatically worse (forces 4 GB upfront allocation)
+#     --min/max-semi-space-size=64     neutral
+#     --single-generation              worse
+#     --separate-gc-phases             worse
+#     --no-concurrent-osr              much worse
+#     --no-parallel-marking            crashes Node 22 (SIGSEGV in ConcurrentMarkerBase)
+#     --predictable-gc-schedule        no reliable effect
+#     --no-concurrent-minor-ms-marking confirmed improvement on Node 22 (see above)
+#
+# =============================================================================
+
+set -e
+cd "$(dirname "$0")"
+
+JEST_BIN="/Users/mdj/Development/OpenSource/reassure/node_modules/jest/bin/jest.js"
+NODE_FLAGS="$*"
+
+mkdir -p .reassure
+
+echo "════════════════════════════════════════════════════════════"
+echo " Node flags: ${NODE_FLAGS:-<none>}"
+echo "════════════════════════════════════════════════════════════"
+echo ""
+
+run_jest() {
+  local label="$1"
+  local output_file="$2"
+  echo "▶ $label …"
+  rm -f "$output_file"
+  # shellcheck disable=SC2086
+  REASSURE_OUTPUT_FILE="$output_file" \
+    REASSURE_SILENT=true \
+    REASSURE_VERBOSE=false \
+    node $NODE_FLAGS "$JEST_BIN" \
+      --runInBand \
+      --testRegex='\.perf\.(tsx?|jsx?)$' \
+      2>&1 | grep -E '(PASS|FAIL|Tests:|✓|✗|×)' | head -20
+  echo ""
+}
+
+run_jest "Run 1/2 (baseline)" ".reassure/baseline.perf"
+run_jest "Run 2/2 (current)"  ".reassure/current.perf"
+
+echo "▶ Stability:"
+node --input-type=commonjs <<'EOF'
+const { compare } = require('@callstack/reassure-compare');
+compare({ outputFormat: 'console' }).catch(e => { console.error(e); process.exit(1); });
+EOF