diff --git a/.Rbuildignore b/.Rbuildignore
index 5b472194..e5e647a4 100644
--- a/.Rbuildignore
+++ b/.Rbuildignore
@@ -33,3 +33,6 @@
 ^.a5c$
 ^README\.html$
 ^tools$
+^tests/testthat/fixtures/tiny-hub-groot/\.db\.cache$
+^tests/testthat/fixtures/tiny-hub-groot/\.ro_attributes$
+^\.worktrees$
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index e3553a59..2f127618 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -40,6 +40,20 @@ Sys.setenv(TESTTHAT_PARALLEL = "TRUE")
 devtools::test()
 ```
 
+A parallel run gives each test file its own copy of the test database under
+`tempdir()`, so point `TMPDIR` at a filesystem with several GB free before
+starting R:
+
+```sh
+TMPDIR=/path/with/space R -e 'Sys.setenv(TESTTHAT_PARALLEL="TRUE"); devtools::test()'
+```
+
+On a shared machine `/tmp` often sits on a full root filesystem. Running out
+of space mid-run leaves partially copied test databases, and the resulting
+failures point at whichever test file happened to use them rather than at the
+disk — expect errors like `Interval test.fixedbin does not exist`. Failures
+that move between files across runs are a symptom of this, not of flaky tests.
+
 GitHub Actions runs `R CMD check` on every push and PR — treat that as the
 authoritative signal.
 
diff --git a/tests/testthat/helper-liftover.R b/tests/testthat/helper-liftover.R
index 71664e5b..471c12d0 100644
--- a/tests/testthat/helper-liftover.R
+++ b/tests/testthat/helper-liftover.R
@@ -108,6 +108,9 @@ setup_db <- function(chrom_defs, return_db = FALSE) {
         {
             unlink(db, recursive = TRUE)
             unlink(fastas)
+            # Do not leave GROOT pointing at the directory we just removed:
+            # the next file in this parallel worker would inherit it.
+            ensure_valid_groot()
         },
         testthat::teardown_env()
     )
diff --git a/tests/testthat/helper-test_db.R b/tests/testthat/helper-test_db.R
index 436fb7d6..122ad951 100644
--- a/tests/testthat/helper-test_db.R
+++ b/tests/testthat/helper-test_db.R
@@ -99,30 +99,82 @@ local_db_state <- function(env = parent.frame()) {
 #' This approach provides complete isolation between parallel test processes
 #' while minimizing disk space and setup time.
 #'
-#' @return Path to the isolated test database
-create_isolated_test_db <- function() {
-    source_db <- if (getOption("gmulticontig.indexed_format", FALSE)) {
+#' Path to the shared, read-only test database
+shared_test_db_path <- function() {
+    if (getOption("gmulticontig.indexed_format", FALSE)) {
         "/net/mraid20/ifs/wisdom/tanay_lab/tgdata/db/tgdb/misha_test_db_indexed/"
     } else {
         "/net/mraid20/export/tgdata/db/tgdb/misha_test_db/"
     }
+}
+
+#' Guarantee this test file leaves a usable GROOT behind
+#'
+#' Files that build their own temporary database and re-root into it leave
+#' .misha$GROOT dangling as soon as that directory is removed (withr deletes
+#' it at the end of the test or the file). Under TESTTHAT_PARALLEL the next
+#' file in the same worker process inherits the dangling root and fails with
+#' something unrelated to its own subject matter - "Database directory does
+#' not exist", "Chromosome chr1 does not exist ... Known chromosomes: chrA",
+#' or "Cannot delete track from read-only database" (a deleted directory is
+#' not writable, so it reads as read-only).
+#'
+#' Call this once at the top of any file that re-roots. It is idempotent and
+#' does nothing when the file leaves a valid root behind.
+ensure_valid_groot <- function() {
+    groot <- if (exists("GROOT", envir = .misha, inherits = FALSE)) {
+        get("GROOT", envir = .misha)
+    } else {
+        NULL
+    }
+    if (is.null(groot) || !dir.exists(groot)) {
+        src <- shared_test_db_path()
+        if (dir.exists(src)) {
+            suppressMessages(gdb.init(src))
+        }
+    }
+    invisible(NULL)
+}
+
+restore_groot_on_exit <- function(envir = parent.frame()) {
+    withr::defer(ensure_valid_groot(), envir = envir)
+}
+
+#' @return Path to the isolated test database
+create_isolated_test_db <- function() {
+    source_db <- shared_test_db_path()
+
+    # Fail loudly instead of leaving a half-built DB behind. A silent failure
+    # here (a full tempdir() is the usual cause) surfaces much later as a
+    # baffling "Interval test.fixedbin does not exist" in an unrelated test.
+    checked_system <- function(cmd) {
+        status <- system(cmd)
+        if (status != 0) {
+            stop(sprintf(
+                "create_isolated_test_db(): `%s` failed with status %d.\nIs tempdir() (%s) out of space? Set TMPDIR to a volume with room.",
+                cmd, status, tempdir()
+            ), call. = FALSE)
+        }
+    }
 
     # Create unique temp dir for this test file/process
     testdb_dir <- tempfile(pattern = "misha_testdb_", tmpdir = tempdir())
     dir.create(testdb_dir, showWarnings = FALSE)
 
     # Copy small files that might be read
-    file.copy(
-        file.path(source_db, "chrom_sizes.txt"),
-        testdb_dir
-    )
+    if (!file.copy(file.path(source_db, "chrom_sizes.txt"), testdb_dir)) {
+        stop(sprintf(
+            "create_isolated_test_db(): failed to copy chrom_sizes.txt into %s (out of space?)",
+            testdb_dir
+        ), call. = FALSE)
+    }
 
     # Copy interval sets and PSSM directories
-    system(sprintf("cp -r %s/intervs %s/", source_db, testdb_dir))
-    system(sprintf("cp -r %s/pssms %s/", source_db, testdb_dir))
+    checked_system(sprintf("cp -r %s/intervs %s/", source_db, testdb_dir))
+    checked_system(sprintf("cp -r %s/pssms %s/", source_db, testdb_dir))
 
     # Symlink the large seq directory (read-only)
-    system(sprintf("ln -s %s/seq %s/seq", source_db, testdb_dir))
+    checked_system(sprintf("ln -s %s/seq %s/seq", source_db, testdb_dir))
 
     # Create tracks directory for this test file
     tracks_dir <- file.path(testdb_dir, "tracks")
@@ -137,7 +189,7 @@ create_isolated_test_db <- function() {
 
     # Symlink each fixture track
     for (track in source_tracks) {
-        system(sprintf(
+        checked_system(sprintf(
             "ln -s %s/tracks/%s %s/tracks/%s",
             source_db, track, testdb_dir, track
         ))
@@ -170,6 +222,14 @@ create_isolated_test_db <- function() {
             if (!is.null(current_groot) && identical(current_groot, testdb_dir)) {
                 if (!is.null(prev_groot) && dir.exists(prev_groot)) {
                     suppressMessages(gdb.init(prev_groot))
+                } else if (dir.exists(source_db)) {
+                    # prev_groot was itself an isolated db that has already been
+                    # torn down. Fall back to the shared test db rather than
+                    # unsetting GROOT: in a parallel run the next file in this
+                    # worker may be one that relies on the ambient root, and it
+                    # would otherwise fail with a confusing "does not exist" or
+                    # "Chromosome chr1 does not exist" error.
+                    suppressMessages(gdb.init(source_db))
                 } else {
                     rm("GROOT", envir = .misha)
                 }
diff --git a/tests/testthat/test-bigset-character-chrom.R b/tests/testthat/test-bigset-character-chrom.R
index b94f967f..c1178d1d 100644
--- a/tests/testthat/test-bigset-character-chrom.R
+++ b/tests/testthat/test-bigset-character-chrom.R
@@ -4,6 +4,8 @@
 # files store $chrom as factor while the persisted .meta zeroline kept the
 # original character class.
 
+restore_groot_on_exit()
+
 test_that("bigset save then load works when input has character chrom", {
     skip_if_not_installed("withr")
     withr::local_options(list(gmulticontig.indexed_format = FALSE))
diff --git a/tests/testthat/test-bigset-fast-load.R b/tests/testthat/test-bigset-fast-load.R
index d3421094..6522ecde 100644
--- a/tests/testthat/test-bigset-fast-load.R
+++ b/tests/testthat/test-bigset-fast-load.R
@@ -1,5 +1,7 @@
 # Tests for fast-path loading of indexed bigsets
 
+restore_groot_on_exit()
+
 test_that(".gintervals.is_indexed_bigset detects indexed bigsets correctly", {
     skip_if_not_installed("withr")
     withr::local_options(list(gmulticontig.indexed_format = FALSE))
diff --git a/tests/testthat/test-db-format-conversion.R b/tests/testthat/test-db-format-conversion.R
index 524b7e91..fb64848b 100644
--- a/tests/testthat/test-db-format-conversion.R
+++ b/tests/testthat/test-db-format-conversion.R
@@ -1,3 +1,5 @@
+restore_groot_on_exit()
+
 load_test_db()
 test_that("gdb.convert_to_indexed converts per-chromosome database to indexed format", {
     local_db_state()
diff --git a/tests/testthat/test-gdb-convert-parallel.R b/tests/testthat/test-gdb-convert-parallel.R
index 40e45de8..64644b92 100644
--- a/tests/testthat/test-gdb-convert-parallel.R
+++ b/tests/testthat/test-gdb-convert-parallel.R
@@ -1,3 +1,5 @@
+restore_groot_on_exit()
+
 load_test_db()
 
 # Tests for parallel gdb.convert_to_indexed (threads argument). Each test
diff --git a/tests/testthat/test-ggenome.R b/tests/testthat/test-ggenome.R
index b49e1676..52121f4c 100644
--- a/tests/testthat/test-ggenome.R
+++ b/tests/testthat/test-ggenome.R
@@ -1,3 +1,5 @@
+create_isolated_test_db()
+
 # Local test helper: read a FASTA file into a named list of sequences
 read_fasta <- function(fasta_path) {
     lines <- readLines(fasta_path)
diff --git a/tests/testthat/test-gintervals-2d-meta-sparse.R b/tests/testthat/test-gintervals-2d-meta-sparse.R
index b11112f7..78119f6d 100644
--- a/tests/testthat/test-gintervals-2d-meta-sparse.R
+++ b/tests/testthat/test-gintervals-2d-meta-sparse.R
@@ -4,6 +4,8 @@
 # load + query path here with a moderate-sized N (1000 contigs) and only a
 # handful of populated chrom-pairs.
 
+restore_groot_on_exit()
+
 build_many_contig_db <- function(n_contigs, defer_envir) {
     skip_if_not_installed("withr")
     tmp_root <- withr::local_tempdir(.local_envir = defer_envir)
diff --git a/tests/testthat/test-gintervals-from-strings.R b/tests/testthat/test-gintervals-from-strings.R
index 584b1c3f..16fe3665 100644
--- a/tests/testthat/test-gintervals-from-strings.R
+++ b/tests/testthat/test-gintervals-from-strings.R
@@ -1,3 +1,5 @@
+create_isolated_test_db()
+
 test_that("gintervals.from_strings parses a single chrom:start-end string", {
     res <- gintervals.from_strings("chr1:100-200")
     expect_equal(nrow(res), 1)
diff --git a/tests/testthat/test-gtrack-create-meta-indexed.R b/tests/testthat/test-gtrack-create-meta-indexed.R
index e3a2cd6e..b4c0788f 100644
--- a/tests/testthat/test-gtrack-create-meta-indexed.R
+++ b/tests/testthat/test-gtrack-create-meta-indexed.R
@@ -13,6 +13,8 @@
 #  2. Meta creation completes for an indexed track where the overwhelming
 #     majority of chromids have no data (the large-contig case).
 
+restore_groot_on_exit()
+
 setup_db <- function(num_chroms, chrom_size = 1e6, indexed = FALSE) {
     tmp_root <- withr::local_tempdir(.local_envir = parent.frame())
     chrom_sizes <- data.frame(
diff --git a/tests/testthat/test-index-cache-invalidation.R b/tests/testthat/test-index-cache-invalidation.R
index e5223a90..4bf3e739 100644
--- a/tests/testthat/test-index-cache-invalidation.R
+++ b/tests/testthat/test-index-cache-invalidation.R
@@ -16,6 +16,8 @@
 # These tests exercise the rm-then-recreate cycle and verify the next
 # read succeeds.
 
+restore_groot_on_exit()
+
 # Build a tiny per-chromosome DB then convert it to indexed format.
 build_indexed_test_db <- function() {
     test_db <- tempfile("misha_idxcache_")
diff --git a/tests/testthat/test-indexed-integration.R b/tests/testthat/test-indexed-integration.R
index 95190da6..a6fb03ab 100644
--- a/tests/testthat/test-indexed-integration.R
+++ b/tests/testthat/test-indexed-integration.R
@@ -1,3 +1,5 @@
+restore_groot_on_exit()
+
 load_test_db()
 # Integration tests for indexed genome format
 
diff --git a/tests/testthat/test-track-indexed-direct.R b/tests/testthat/test-track-indexed-direct.R
index b4633988..1d31bf8d 100644
--- a/tests/testthat/test-track-indexed-direct.R
+++ b/tests/testthat/test-track-indexed-direct.R
@@ -4,6 +4,8 @@
 
 # Build a small indexed DB from a 5-chromosome FASTA and create the
 # named track via gtrack.create. Returns the path to the track dir.
+restore_groot_on_exit()
+
 .make_indexed_db_with_track <- function(track_name, bin_size = 4) {
     test_fasta <- tempfile(fileext = ".fasta")
     cat(">chr1\nACTGACTGACTGACTGACTGACTGACTGACTG\n",