diff --git a/tests/testthat/test-ts-ratchet-asan-oob.R b/tests/testthat/test-ts-ratchet-asan-oob.R
new file mode 100644
index 000000000..e3730bc8f
--- /dev/null
+++ b/tests/testthat/test-ts-ratchet-asan-oob.R
@@ -0,0 +1,66 @@
+# Memory-safety stress guard for the C++ ratchet on multistate data.
+# Deliberately NOT gated by skip_extended(), so the gcc-ASAN workflow
+# (ASan.yml runs plain `tests`) executes it: many repeated ts_ratchet_search
+# calls in one process on a 30-tip x 25-char 3-state matrix, which is exactly
+# the shape that would surface a heap out-of-bounds write in the TBR/fitch
+# undo/rescore machinery if one were ever introduced.
+#
+# Origin: a 2026-07-16 report of an exit-139 heap-corruption crash here
+# (memory node: ratchet-multistate-segfault). Investigation could NOT
+# reproduce it on ANY clean build — Linux gcc-ASAN clean, and a fresh local
+# Windows -O2 build ran 500 iterations without a crash — while the search
+# kernels were byte-identical to the reported commit. The crash was most
+# consistent with a stale-object / mixed-ABI incremental build (see memory
+# node stale-object-abi-gotcha), not a defect in committed source.
+#
+# So: this test is EXPECTED TO PASS. If it ever fails, first REBUILD CLEAN
+# (rm src/*.o; CCACHE_DISABLE=1 R CMD INSTALL --preclean) and re-run before
+# treating it as a genuine source regression.
+
+# --- Config 1 (PRIMARY): ZERO_ONLY perturbation, 3-state EW -----------------
+# 30 tips x 25 chars, states 0:2, equal weights, perturbMode = 0 (ZERO_ONLY).
+# This is the config that crashes soonest in the documented repro.
+
+test_that("ratchet ZERO_ONLY 3-state EW survives repeated calls (ASan)", {
+  set.seed(6284)
+  m <- matrix(sample(0:2, 30 * 25, replace = TRUE), nrow = 30,
+              dimnames = list(paste0("t", 1:30), NULL))
+  dataset <- MatrixToPhyDat(m)
+  ds <- make_ts_data(dataset)
+  tr <- as.phylo(1, 30)
+
+  ok <- TRUE
+  for (i in 1:30) {
+    set.seed(i)
+    r <- TreeSearch:::ts_ratchet_search(
+      tr$edge, ds$contrast, ds$tip_data, ds$weight, ds$levels,
+      nCycles = 5L, perturbMode = 0L, perturbProb = 0.15)
+    ok <- ok && is.finite(r$score) && r$score >= 0
+  }
+  expect_true(ok)
+})
+
+# --- Config 2 (SECONDARY): UPWEIGHT_ONLY perturbation, IW (concavity 3) ------
+# perturbMode = 1 with min_steps + concavity engages the IW pattern_freq
+# upweight path. Only reached if Config 1 is clean (ASan halts on first hit).
+
+test_that("ratchet UPWEIGHT_ONLY IW survives repeated calls (ASan)", {
+  set.seed(6284)
+  m <- matrix(sample(0:2, 30 * 25, replace = TRUE), nrow = 30,
+              dimnames = list(paste0("t", 1:30), NULL))
+  dataset <- MatrixToPhyDat(m)
+  ds <- make_ts_data(dataset)
+  minSteps <- as.integer(MinimumLength(dataset, compress = TRUE))
+  tr <- as.phylo(1, 30)
+
+  ok <- TRUE
+  for (i in 1:30) {
+    set.seed(i)
+    r <- TreeSearch:::ts_ratchet_search(
+      tr$edge, ds$contrast, ds$tip_data, ds$weight, ds$levels,
+      nCycles = 5L, perturbMode = 1L, perturbProb = 0.15,
+      min_steps = minSteps, concavity = 3.0)
+    ok <- ok && is.finite(r$score) && r$score >= 0
+  }
+  expect_true(ok)
+})