diff --git a/CHANGELOG.md b/CHANGELOG.md
index ad3f0a9b..c39f90b1 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -13,6 +13,7 @@
 - `--jobs auto` / `-j auto` caps parallel concurrency at the CPU core count (portable across Linux/macOS/BSD); the default stays unlimited (#766)
 
 ### Changed
+- Faster cold start: the deferred-output scratch files (failures/skipped/incomplete/risky/profile/rerun) now share one run-unique directory instead of forking `mktemp` once each (6 -> 0 `mktemp` forks per cold start; ~62ms -> ~50ms per nested run on bash 3.2). No behaviour change (#798)
 - Faster cold start: `check_os` detects the OS with a single `uname` fork at load instead of two (matters across the acceptance suite's nested `bashunit` runs). No behaviour change (#798)
 - Faster test execution: removed the remaining round-2 fork leftovers — `wait_for_job_slot` counts running background jobs without forking `wc` on every poll (145 -> 0 `wc` forks across a small parallel run), and `check_duplicate_functions` folds its per-file grep+awk+sort+uniq scan into a single awk pass. No behaviour change (#761)
 - `assert_equals`/`assert_same` failures with multiline values now render a git word-diff below the header (requires git, opt out with `BASHUNIT_NO_DIFF=true`, respects `--no-color`); machine reports keep the raw values (#777)
diff --git a/src/env.sh b/src/env.sh
index 426af435..a1a9c22a 100644
--- a/src/env.sh
+++ b/src/env.sh
@@ -486,14 +486,23 @@ TERMINAL_WIDTH="$(bashunit::env::find_terminal_width)"
 CAT="$(command -v cat)"
 GREP="$(command -v grep)"
 MKTEMP="$(command -v mktemp)"
-FAILURES_OUTPUT_PATH=$("$MKTEMP")
-SKIPPED_OUTPUT_PATH=$("$MKTEMP")
-INCOMPLETE_OUTPUT_PATH=$("$MKTEMP")
-RISKY_OUTPUT_PATH=$("$MKTEMP")
-PROFILE_OUTPUT_PATH=$("$MKTEMP")
+# Deferred-output scratch files. Each used to be its own `mktemp` fork; at ~258
+# nested cold starts in the acceptance suite that is ~1.5k forks and dominates
+# cold-start cost (#798). Derive them from one run-unique directory instead:
+# `bashunit::random_str` is fork-free and every consumer appends with `>>` (which
+# creates the file lazily) or guards reads with `[ -s ... ]`, so the files need
+# not be pre-created. The random suffix keeps the directory unique across
+# recursive and parallel invocations, matching TEMP_DIR_PARALLEL_TEST_SUITE.
+_BASHUNIT_RUN_OUTPUT_DIR="${TMPDIR:-/tmp}/bashunit/run/${_BASHUNIT_OS:-Unknown}/$(bashunit::random_str 8)"
+mkdir -p "$_BASHUNIT_RUN_OUTPUT_DIR" 2>/dev/null || true
+FAILURES_OUTPUT_PATH="$_BASHUNIT_RUN_OUTPUT_DIR/failures"
+SKIPPED_OUTPUT_PATH="$_BASHUNIT_RUN_OUTPUT_DIR/skipped"
+INCOMPLETE_OUTPUT_PATH="$_BASHUNIT_RUN_OUTPUT_DIR/incomplete"
+RISKY_OUTPUT_PATH="$_BASHUNIT_RUN_OUTPUT_DIR/risky"
+PROFILE_OUTPUT_PATH="$_BASHUNIT_RUN_OUTPUT_DIR/profile"
 # Collects "<test_file>:<function_name>" for every failing test in a run so the
 # next --rerun-failed can replay just those. Shared across parallel subshells.
-RERUN_FAILED_OUTPUT_PATH=$("$MKTEMP")
+RERUN_FAILED_OUTPUT_PATH="$_BASHUNIT_RUN_OUTPUT_DIR/rerun-failed"
 
 # Initialize temp directory once at startup for performance
 BASHUNIT_TEMP_DIR="${TMPDIR:-/tmp}/bashunit/tmp"
diff --git a/tests/acceptance/bashunit_coldstart_forks_test.sh b/tests/acceptance/bashunit_coldstart_forks_test.sh
new file mode 100644
index 00000000..470442e8
--- /dev/null
+++ b/tests/acceptance/bashunit_coldstart_forks_test.sh
@@ -0,0 +1,25 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Regression guard for #798. bashunit's cold start historically forked `mktemp`
+# once per deferred-output scratch file (failures, skipped, incomplete, risky,
+# profile, rerun). Across the acceptance suite's ~258 nested runs that is ~1.5k
+# forks and dominates wall-clock. Those scratch files must now be derived from a
+# single run directory (no per-file fork), so a cold start forks `mktemp` zero
+# times for them.
+function test_coldstart_does_not_fork_mktemp_per_scratch_file() {
+  if bashunit::check_os::is_windows; then
+    bashunit::skip "process tracing is unreliable under Git Bash" && return
+  fi
+
+  local trace
+  trace="$(PS4='+ ' bash -x ./bashunit --version 2>&1 >/dev/null)"
+
+  # Count real `mktemp` process executions: a traced line whose command token is
+  # an absolute path ending in `mktemp` with no arguments. This excludes the
+  # `command -v mktemp` probe and the `MKTEMP=` variable assignment.
+  local mktemp_forks
+  mktemp_forks="$(printf '%s\n' "$trace" | grep -cE '^\++ +/[^ ]*mktemp$' || true)"
+
+  assert_less_than 1 "$mktemp_forks"
+}