fix(runner): capture parallel worker stderr instead of discarding it

CHANGELOG.md

@@ -16,6 +16,7 @@
 - The `--parallel` unsupported-OS warning no longer claims Alpine is excluded: Alpine has been a supported parallel platform since the race conditions were fixed, the message was simply never updated
 
 ### Fixed
+- `--parallel` no longer discards stderr written inside a worker but outside a test body, which made the same run report differently depending on the mode. Each file worker had been spawned with `2>/dev/null` since #358 to keep its noise off the progress line; its stderr is now captured per file and rendered afterwards as a `Stderr from <file>` block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Output written by a test *body* was never affected — it is merged into that test's captured stdout and already appeared in its failure block (#864)
 - The minimum-bash gate now compares the minor version as well as the major, so it enforces whatever `BASHUNIT_MIN_BASH_VERSION` declares. It previously accepted any `3.x` regardless of the stated minimum, and a version string carrying a suffix (`5.2.37(1)-release`) is now parsed instead of tripping the comparison. The floor itself is unchanged at **Bash 3.0+**; `printf -v`, `+=` and `[[ =~ ]]` are now rejected in `src/` by the compatibility gate, since 3.0/3.1 lack the first two and 3.2 changed quoted-pattern semantics for the third
 - An empty entry in `.env` no longer overrides a value the caller exported or set on the command line. `.env` is sourced under `set -o allexport`, so every line was an unconditional assignment: merely *listing* a name blanked it, and `BASHUNIT_OUTPUT_FORMAT=tap ./bashunit` silently stopped working in any project whose `.env` mentioned that setting. An empty entry now means "not configured here"; an entry with a value still takes effect. This had been actively concealing defects — two of the bugs fixed in #879 were not reproducible from inside a repo checkout for exactly this reason (#865)
 - A malformed benchmark annotation is now an error instead of a silent fallback to the default: `@revs=abc` quietly ran a single revolution, `@its=abc` a single iteration, and `@max_ms=abc` dropped the threshold entirely so the benchmark could never fail — each reporting success while measuring something other than what was written (#884)

src/console_results.sh

@@ -559,6 +559,24 @@ function bashunit::console_results::print_error_test() {
  bashunit::state::print_line "error" "$line"
 }
 
+##
+# Render stderr a parallel worker wrote outside any test body.
+# A sequential run lets this straight through to the terminal; parallel workers
+# have it captured per file so concurrent writes cannot shred the progress
+# line. Attributed to the file, not a test: it is emitted where no test owns it
+# (data providers, hook plumbing). Test-body stderr is merged into the captured
+# stdout and still surfaces in that test's own failure block.
+# Arguments: $1 - test file the worker ran, $2 - captured stderr file
+##
+function bashunit::console_results::print_worker_stderr() {
+ local test_file="$1"
+ local stderr_file="$2"
+
+ printf "\n%sStderr from %s%s\n" \
+ "$_BASHUNIT_COLOR_SKIPPED" "$test_file" "$_BASHUNIT_COLOR_DEFAULT"
+ sed 's/^/|/' "$stderr_file"
+}
+
 function bashunit::console_results::print_failing_tests_and_reset() {
  if [ -s "$FAILURES_OUTPUT_PATH" ]; then
  local total_failed

src/env.sh

@@ -654,6 +654,11 @@ 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"
+# Prefix for one file-per-worker capture of stderr written inside a parallel
+# worker but outside a test body. An ordinal is appended per spawned worker;
+# these must not live under TEMP_DIR_PARALLEL_TEST_SUITE, whose every entry is
+# walked by state::aggregate_parallel_results (#864).
+WORKER_STDERR_OUTPUT_PREFIX="$_BASHUNIT_RUN_OUTPUT_DIR/worker-stderr"
 # Collects "<test_file>:<function_name>" for every failing test in a run so the
 # next --rerun-failed can replay just those. Shared across parallel subshells.
 RERUN_FAILED_OUTPUT_PATH="$_BASHUNIT_RUN_OUTPUT_DIR/rerun-failed"

src/runner.sh

@@ -396,6 +396,9 @@ function bashunit::runner::load_test_files() {
  files=("$@")
  local -a scripts_ids=()
  local scripts_ids_count=0
+ local -a worker_stderr_paths=()
+ local -a worker_stderr_owners=()
+ local worker_stderr_count=0
 
  # Randomize file execution order (deterministic for the resolved seed).
  if bashunit::env::is_random_order_enabled; then
@@ -538,7 +541,15 @@ function bashunit::runner::load_test_files() {
  local _cached_fns="$functions_for_script"
  if bashunit::parallel::is_enabled; then
  bashunit::runner::wait_for_job_slot
- bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>/dev/null &
+ # Capture rather than discard: a worker's stderr cannot be written
+ # straight to the terminal without shredding the progress line, but
+ # dropping it made the same run report differently under --parallel
+ # (#358 added the discard, #864 replaced it with this capture).
+ local _worker_stderr="${WORKER_STDERR_OUTPUT_PREFIX}.${worker_stderr_count}"
+ worker_stderr_paths[worker_stderr_count]="$_worker_stderr"
+ worker_stderr_owners[worker_stderr_count]="$test_file"
+ worker_stderr_count=$((worker_stderr_count + 1))
+ bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>"$_worker_stderr" &
  else
  bashunit::runner::call_test_functions "$test_file" "$_cached_fns"
  fi
@@ -561,6 +572,16 @@ function bashunit::runner::load_test_files() {
  disown "$spinner_pid" 2>/dev/null || true
  kill "$spinner_pid" 2>/dev/null || true
  printf "\r \r" # Clear the spinner output
+
+ local _stderr_idx=0
+ while [ "$_stderr_idx" -lt "$worker_stderr_count" ]; do
+ if [ -s "${worker_stderr_paths[_stderr_idx]:-}" ]; then
+ bashunit::console_results::print_worker_stderr \
+ "${worker_stderr_owners[_stderr_idx]:-}" "${worker_stderr_paths[_stderr_idx]:-}"
+ fi
+ _stderr_idx=$((_stderr_idx + 1))
+ done
+
  local script_id
  for script_id in "${scripts_ids[@]+"${scripts_ids[@]}"}"; do
  export BASHUNIT_CURRENT_SCRIPT_ID="${script_id}"

tests/acceptance/fixtures/test_worker_stderr.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+
+# A data provider runs inside the parallel file worker but outside any test
+# body, so its stderr is the plain case that the worker-level redirect used to
+# discard. Test-body stderr is a different path: it is merged into the captured
+# stdout and already surfaces in the failure block.
+
+function data_provider_noisy() {
+ echo "WORKER-SCOPE-DIAGNOSTIC" >&2
+ echo "1"
+}
+
+# @data_provider data_provider_noisy
+function test_uses_a_noisy_provider() {
+ assert_equals "1" "$1"
+}

tests/acceptance/worker_stderr_test.sh

@@ -0,0 +1,42 @@
+#!/usr/bin/env bash
+
+# Parallel workers used to run with `2>/dev/null`, so anything written to
+# stderr inside the worker but outside a test body vanished, and the same run
+# reported differently depending on --parallel (#864).
+#
+# The fixture's provider is also executed by the test-counting pass in the main
+# shell, so merely *finding* the diagnostic proves nothing — it leaks in from
+# that pass either way. What the bug actually broke is the two modes agreeing,
+# so that is what these compare.
+
+FIXTURE="tests/acceptance/fixtures/test_worker_stderr.sh"
+
+function _count_diagnostic_in() {
+ local parallel_flag="$1"
+
+ ./bashunit "$parallel_flag" "$FIXTURE" 2>&1 |
+ grep -c "WORKER-SCOPE-DIAGNOSTIC"
+}
+
+function test_parallel_reports_worker_stderr_as_often_as_sequential() {
+ local sequential parallel
+ sequential="$(_count_diagnostic_in --no-parallel)"
+ parallel="$(_count_diagnostic_in --parallel)"
+
+ assert_not_equals "0" "$sequential"
+ assert_equals "$sequential" "$parallel"
+}
+
+function test_parallel_run_attributes_worker_stderr_to_its_file() {
+ local output
+ output=$(./bashunit --parallel "$FIXTURE" 2>&1)
+
+ assert_contains "Stderr from $FIXTURE" "$output"
+}
+
+function test_parallel_run_stays_quiet_when_no_worker_writes_stderr() {
+ local output
+ output=$(./bashunit --parallel tests/acceptance/fixtures/test_coverage_engine.sh 2>&1)
+
+ assert_not_contains "Stderr from" "$output"
+}

CHANGELOG.md

@@ -16,6 +16,7 @@
 - The `--parallel` unsupported-OS warning no longer claims Alpine is excluded: Alpine has been a supported parallel platform since the race conditions were fixed, the message was simply never updated
 
 ### Fixed
+- `--parallel` no longer discards stderr written inside a worker but outside a test body, which made the same run report differently depending on the mode. Each file worker had been spawned with `2>/dev/null` since #358 to keep its noise off the progress line; its stderr is now captured per file and rendered afterwards as a `Stderr from <file>` block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Output written by a test *body* was never affected — it is merged into that test's captured stdout and already appeared in its failure block (#864)
 - The minimum-bash gate now compares the minor version as well as the major, so it enforces whatever `BASHUNIT_MIN_BASH_VERSION` declares. It previously accepted any `3.x` regardless of the stated minimum, and a version string carrying a suffix (`5.2.37(1)-release`) is now parsed instead of tripping the comparison. The floor itself is unchanged at **Bash 3.0+**; `printf -v`, `+=` and `[[ =~ ]]` are now rejected in `src/` by the compatibility gate, since 3.0/3.1 lack the first two and 3.2 changed quoted-pattern semantics for the third
 - An empty entry in `.env` no longer overrides a value the caller exported or set on the command line. `.env` is sourced under `set -o allexport`, so every line was an unconditional assignment: merely *listing* a name blanked it, and `BASHUNIT_OUTPUT_FORMAT=tap ./bashunit` silently stopped working in any project whose `.env` mentioned that setting. An empty entry now means "not configured here"; an entry with a value still takes effect. This had been actively concealing defects — two of the bugs fixed in #879 were not reproducible from inside a repo checkout for exactly this reason (#865)
 - A malformed benchmark annotation is now an error instead of a silent fallback to the default: `@revs=abc` quietly ran a single revolution, `@its=abc` a single iteration, and `@max_ms=abc` dropped the threshold entirely so the benchmark could never fail — each reporting success while measuring something other than what was written (#884)

src/console_results.sh

@@ -559,6 +559,24 @@ function bashunit::console_results::print_error_test() {
  bashunit::state::print_line "error" "$line"
 }
 
+##
+# Render stderr a parallel worker wrote outside any test body.
+# A sequential run lets this straight through to the terminal; parallel workers
+# have it captured per file so concurrent writes cannot shred the progress
+# line. Attributed to the file, not a test: it is emitted where no test owns it
+# (data providers, hook plumbing). Test-body stderr is merged into the captured
+# stdout and still surfaces in that test's own failure block.
+# Arguments: $1 - test file the worker ran, $2 - captured stderr file
+##
+function bashunit::console_results::print_worker_stderr() {
+ local test_file="$1"
+ local stderr_file="$2"
+
+ printf "\n%sStderr from %s%s\n" \
+ "$_BASHUNIT_COLOR_SKIPPED" "$test_file" "$_BASHUNIT_COLOR_DEFAULT"
+ sed 's/^/|/' "$stderr_file"
+}
+
 function bashunit::console_results::print_failing_tests_and_reset() {
  if [ -s "$FAILURES_OUTPUT_PATH" ]; then
  local total_failed

src/env.sh

@@ -654,6 +654,11 @@ 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"
+# Prefix for one file-per-worker capture of stderr written inside a parallel
+# worker but outside a test body. An ordinal is appended per spawned worker;
+# these must not live under TEMP_DIR_PARALLEL_TEST_SUITE, whose every entry is
+# walked by state::aggregate_parallel_results (#864).
+WORKER_STDERR_OUTPUT_PREFIX="$_BASHUNIT_RUN_OUTPUT_DIR/worker-stderr"
 # Collects "<test_file>:<function_name>" for every failing test in a run so the
 # next --rerun-failed can replay just those. Shared across parallel subshells.
 RERUN_FAILED_OUTPUT_PATH="$_BASHUNIT_RUN_OUTPUT_DIR/rerun-failed"

src/runner.sh

@@ -396,6 +396,9 @@ function bashunit::runner::load_test_files() {
  files=("$@")
  local -a scripts_ids=()
  local scripts_ids_count=0
+ local -a worker_stderr_paths=()
+ local -a worker_stderr_owners=()
+ local worker_stderr_count=0
 
  # Randomize file execution order (deterministic for the resolved seed).
  if bashunit::env::is_random_order_enabled; then
@@ -538,7 +541,15 @@ function bashunit::runner::load_test_files() {
  local _cached_fns="$functions_for_script"
  if bashunit::parallel::is_enabled; then
  bashunit::runner::wait_for_job_slot
- bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>/dev/null &
+ # Capture rather than discard: a worker's stderr cannot be written
+ # straight to the terminal without shredding the progress line, but
+ # dropping it made the same run report differently under --parallel
+ # (#358 added the discard, #864 replaced it with this capture).
+ local _worker_stderr="${WORKER_STDERR_OUTPUT_PREFIX}.${worker_stderr_count}"
+ worker_stderr_paths[worker_stderr_count]="$_worker_stderr"
+ worker_stderr_owners[worker_stderr_count]="$test_file"
+ worker_stderr_count=$((worker_stderr_count + 1))
+ bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>"$_worker_stderr" &
  else
  bashunit::runner::call_test_functions "$test_file" "$_cached_fns"
  fi
@@ -561,6 +572,16 @@ function bashunit::runner::load_test_files() {
  disown "$spinner_pid" 2>/dev/null || true
  kill "$spinner_pid" 2>/dev/null || true
  printf "\r \r" # Clear the spinner output
+
+ local _stderr_idx=0
+ while [ "$_stderr_idx" -lt "$worker_stderr_count" ]; do
+ if [ -s "${worker_stderr_paths[_stderr_idx]:-}" ]; then
+ bashunit::console_results::print_worker_stderr \
+ "${worker_stderr_owners[_stderr_idx]:-}" "${worker_stderr_paths[_stderr_idx]:-}"
+ fi
+ _stderr_idx=$((_stderr_idx + 1))
+ done
+
  local script_id
  for script_id in "${scripts_ids[@]+"${scripts_ids[@]}"}"; do
  export BASHUNIT_CURRENT_SCRIPT_ID="${script_id}"

tests/acceptance/fixtures/test_worker_stderr.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+
+# A data provider runs inside the parallel file worker but outside any test
+# body, so its stderr is the plain case that the worker-level redirect used to
+# discard. Test-body stderr is a different path: it is merged into the captured
+# stdout and already surfaces in the failure block.
+
+function data_provider_noisy() {
+ echo "WORKER-SCOPE-DIAGNOSTIC" >&2
+ echo "1"
+}
+
+# @data_provider data_provider_noisy
+function test_uses_a_noisy_provider() {
+ assert_equals "1" "$1"
+}

tests/acceptance/worker_stderr_test.sh

@@ -0,0 +1,42 @@
+#!/usr/bin/env bash
+
+# Parallel workers used to run with `2>/dev/null`, so anything written to
+# stderr inside the worker but outside a test body vanished, and the same run
+# reported differently depending on --parallel (#864).
+#
+# The fixture's provider is also executed by the test-counting pass in the main
+# shell, so merely *finding* the diagnostic proves nothing — it leaks in from
+# that pass either way. What the bug actually broke is the two modes agreeing,
+# so that is what these compare.
+
+FIXTURE="tests/acceptance/fixtures/test_worker_stderr.sh"
+
+function _count_diagnostic_in() {
+ local parallel_flag="$1"
+
+ ./bashunit "$parallel_flag" "$FIXTURE" 2>&1 |
+ grep -c "WORKER-SCOPE-DIAGNOSTIC"
+}
+
+function test_parallel_reports_worker_stderr_as_often_as_sequential() {
+ local sequential parallel
+ sequential="$(_count_diagnostic_in --no-parallel)"
+ parallel="$(_count_diagnostic_in --parallel)"
+
+ assert_not_equals "0" "$sequential"
+ assert_equals "$sequential" "$parallel"
+}
+
+function test_parallel_run_attributes_worker_stderr_to_its_file() {
+ local output
+ output=$(./bashunit --parallel "$FIXTURE" 2>&1)
+
+ assert_contains "Stderr from $FIXTURE" "$output"
+}
+
+function test_parallel_run_stays_quiet_when_no_worker_writes_stderr() {
+ local output
+ output=$(./bashunit --parallel tests/acceptance/fixtures/test_coverage_engine.sh 2>&1)
+
+ assert_not_contains "Stderr from" "$output"
+}

CHANGELOG.md

@@ -16,6 +16,7 @@
 - The `--parallel` unsupported-OS warning no longer claims Alpine is excluded: Alpine has been a supported parallel platform since the race conditions were fixed, the message was simply never updated
 
 ### Fixed
+- `--parallel` no longer discards stderr written inside a worker but outside a test body, which made the same run report differently depending on the mode. Each file worker had been spawned with `2>/dev/null` since #358 to keep its noise off the progress line; its stderr is now captured per file and rendered afterwards as a `Stderr from <file>` block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Output written by a test *body* was never affected — it is merged into that test's captured stdout and already appeared in its failure block (#864)
 - The minimum-bash gate now compares the minor version as well as the major, so it enforces whatever `BASHUNIT_MIN_BASH_VERSION` declares. It previously accepted any `3.x` regardless of the stated minimum, and a version string carrying a suffix (`5.2.37(1)-release`) is now parsed instead of tripping the comparison. The floor itself is unchanged at **Bash 3.0+**; `printf -v`, `+=` and `[[ =~ ]]` are now rejected in `src/` by the compatibility gate, since 3.0/3.1 lack the first two and 3.2 changed quoted-pattern semantics for the third
 - An empty entry in `.env` no longer overrides a value the caller exported or set on the command line. `.env` is sourced under `set -o allexport`, so every line was an unconditional assignment: merely *listing* a name blanked it, and `BASHUNIT_OUTPUT_FORMAT=tap ./bashunit` silently stopped working in any project whose `.env` mentioned that setting. An empty entry now means "not configured here"; an entry with a value still takes effect. This had been actively concealing defects — two of the bugs fixed in #879 were not reproducible from inside a repo checkout for exactly this reason (#865)
 - A malformed benchmark annotation is now an error instead of a silent fallback to the default: `@revs=abc` quietly ran a single revolution, `@its=abc` a single iteration, and `@max_ms=abc` dropped the threshold entirely so the benchmark could never fail — each reporting success while measuring something other than what was written (#884)

src/console_results.sh

@@ -559,6 +559,24 @@ function bashunit::console_results::print_error_test() {
  bashunit::state::print_line "error" "$line"
 }
 
+##
+# Render stderr a parallel worker wrote outside any test body.
+# A sequential run lets this straight through to the terminal; parallel workers
+# have it captured per file so concurrent writes cannot shred the progress
+# line. Attributed to the file, not a test: it is emitted where no test owns it
+# (data providers, hook plumbing). Test-body stderr is merged into the captured
+# stdout and still surfaces in that test's own failure block.
+# Arguments: $1 - test file the worker ran, $2 - captured stderr file
+##
+function bashunit::console_results::print_worker_stderr() {
+ local test_file="$1"
+ local stderr_file="$2"
+
+ printf "\n%sStderr from %s%s\n" \
+ "$_BASHUNIT_COLOR_SKIPPED" "$test_file" "$_BASHUNIT_COLOR_DEFAULT"
+ sed 's/^/|/' "$stderr_file"
+}
+
 function bashunit::console_results::print_failing_tests_and_reset() {
  if [ -s "$FAILURES_OUTPUT_PATH" ]; then
  local total_failed

src/env.sh

@@ -654,6 +654,11 @@ 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"
+# Prefix for one file-per-worker capture of stderr written inside a parallel
+# worker but outside a test body. An ordinal is appended per spawned worker;
+# these must not live under TEMP_DIR_PARALLEL_TEST_SUITE, whose every entry is
+# walked by state::aggregate_parallel_results (#864).
+WORKER_STDERR_OUTPUT_PREFIX="$_BASHUNIT_RUN_OUTPUT_DIR/worker-stderr"
 # Collects "<test_file>:<function_name>" for every failing test in a run so the
 # next --rerun-failed can replay just those. Shared across parallel subshells.
 RERUN_FAILED_OUTPUT_PATH="$_BASHUNIT_RUN_OUTPUT_DIR/rerun-failed"

src/runner.sh

@@ -396,6 +396,9 @@ function bashunit::runner::load_test_files() {
  files=("$@")
  local -a scripts_ids=()
  local scripts_ids_count=0
+ local -a worker_stderr_paths=()
+ local -a worker_stderr_owners=()
+ local worker_stderr_count=0
 
  # Randomize file execution order (deterministic for the resolved seed).
  if bashunit::env::is_random_order_enabled; then
@@ -538,7 +541,15 @@ function bashunit::runner::load_test_files() {
  local _cached_fns="$functions_for_script"
  if bashunit::parallel::is_enabled; then
  bashunit::runner::wait_for_job_slot
- bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>/dev/null &
+ # Capture rather than discard: a worker's stderr cannot be written
+ # straight to the terminal without shredding the progress line, but
+ # dropping it made the same run report differently under --parallel
+ # (#358 added the discard, #864 replaced it with this capture).
+ local _worker_stderr="${WORKER_STDERR_OUTPUT_PREFIX}.${worker_stderr_count}"
+ worker_stderr_paths[worker_stderr_count]="$_worker_stderr"
+ worker_stderr_owners[worker_stderr_count]="$test_file"
+ worker_stderr_count=$((worker_stderr_count + 1))
+ bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>"$_worker_stderr" &
  else
  bashunit::runner::call_test_functions "$test_file" "$_cached_fns"
  fi
@@ -561,6 +572,16 @@ function bashunit::runner::load_test_files() {
  disown "$spinner_pid" 2>/dev/null || true
  kill "$spinner_pid" 2>/dev/null || true
  printf "\r \r" # Clear the spinner output
+
+ local _stderr_idx=0
+ while [ "$_stderr_idx" -lt "$worker_stderr_count" ]; do
+ if [ -s "${worker_stderr_paths[_stderr_idx]:-}" ]; then
+ bashunit::console_results::print_worker_stderr \
+ "${worker_stderr_owners[_stderr_idx]:-}" "${worker_stderr_paths[_stderr_idx]:-}"
+ fi
+ _stderr_idx=$((_stderr_idx + 1))
+ done
+
  local script_id
  for script_id in "${scripts_ids[@]+"${scripts_ids[@]}"}"; do
  export BASHUNIT_CURRENT_SCRIPT_ID="${script_id}"

tests/acceptance/fixtures/test_worker_stderr.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+
+# A data provider runs inside the parallel file worker but outside any test
+# body, so its stderr is the plain case that the worker-level redirect used to
+# discard. Test-body stderr is a different path: it is merged into the captured
+# stdout and already surfaces in the failure block.
+
+function data_provider_noisy() {
+ echo "WORKER-SCOPE-DIAGNOSTIC" >&2
+ echo "1"
+}
+
+# @data_provider data_provider_noisy
+function test_uses_a_noisy_provider() {
+ assert_equals "1" "$1"
+}

tests/acceptance/worker_stderr_test.sh

@@ -0,0 +1,42 @@
+#!/usr/bin/env bash
+
+# Parallel workers used to run with `2>/dev/null`, so anything written to
+# stderr inside the worker but outside a test body vanished, and the same run
+# reported differently depending on --parallel (#864).
+#
+# The fixture's provider is also executed by the test-counting pass in the main
+# shell, so merely *finding* the diagnostic proves nothing — it leaks in from
+# that pass either way. What the bug actually broke is the two modes agreeing,
+# so that is what these compare.
+
+FIXTURE="tests/acceptance/fixtures/test_worker_stderr.sh"
+
+function _count_diagnostic_in() {
+ local parallel_flag="$1"
+
+ ./bashunit "$parallel_flag" "$FIXTURE" 2>&1 |
+ grep -c "WORKER-SCOPE-DIAGNOSTIC"
+}
+
+function test_parallel_reports_worker_stderr_as_often_as_sequential() {
+ local sequential parallel
+ sequential="$(_count_diagnostic_in --no-parallel)"
+ parallel="$(_count_diagnostic_in --parallel)"
+
+ assert_not_equals "0" "$sequential"
+ assert_equals "$sequential" "$parallel"
+}
+
+function test_parallel_run_attributes_worker_stderr_to_its_file() {
+ local output
+ output=$(./bashunit --parallel "$FIXTURE" 2>&1)
+
+ assert_contains "Stderr from $FIXTURE" "$output"
+}
+
+function test_parallel_run_stays_quiet_when_no_worker_writes_stderr() {
+ local output
+ output=$(./bashunit --parallel tests/acceptance/fixtures/test_coverage_engine.sh 2>&1)
+
+ assert_not_contains "Stderr from" "$output"
+}

CHANGELOG.md

@@ -16,6 +16,7 @@
 - The `--parallel` unsupported-OS warning no longer claims Alpine is excluded: Alpine has been a supported parallel platform since the race conditions were fixed, the message was simply never updated
 
 ### Fixed
+- `--parallel` no longer discards stderr written inside a worker but outside a test body, which made the same run report differently depending on the mode. Each file worker had been spawned with `2>/dev/null` since #358 to keep its noise off the progress line; its stderr is now captured per file and rendered afterwards as a `Stderr from <file>` block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Output written by a test *body* was never affected — it is merged into that test's captured stdout and already appeared in its failure block (#864)
 - The minimum-bash gate now compares the minor version as well as the major, so it enforces whatever `BASHUNIT_MIN_BASH_VERSION` declares. It previously accepted any `3.x` regardless of the stated minimum, and a version string carrying a suffix (`5.2.37(1)-release`) is now parsed instead of tripping the comparison. The floor itself is unchanged at **Bash 3.0+**; `printf -v`, `+=` and `[[ =~ ]]` are now rejected in `src/` by the compatibility gate, since 3.0/3.1 lack the first two and 3.2 changed quoted-pattern semantics for the third
 - An empty entry in `.env` no longer overrides a value the caller exported or set on the command line. `.env` is sourced under `set -o allexport`, so every line was an unconditional assignment: merely *listing* a name blanked it, and `BASHUNIT_OUTPUT_FORMAT=tap ./bashunit` silently stopped working in any project whose `.env` mentioned that setting. An empty entry now means "not configured here"; an entry with a value still takes effect. This had been actively concealing defects — two of the bugs fixed in #879 were not reproducible from inside a repo checkout for exactly this reason (#865)
 - A malformed benchmark annotation is now an error instead of a silent fallback to the default: `@revs=abc` quietly ran a single revolution, `@its=abc` a single iteration, and `@max_ms=abc` dropped the threshold entirely so the benchmark could never fail — each reporting success while measuring something other than what was written (#884)

src/console_results.sh

@@ -559,6 +559,24 @@ function bashunit::console_results::print_error_test() {
  bashunit::state::print_line "error" "$line"
 }
 
+##
+# Render stderr a parallel worker wrote outside any test body.
+# A sequential run lets this straight through to the terminal; parallel workers
+# have it captured per file so concurrent writes cannot shred the progress
+# line. Attributed to the file, not a test: it is emitted where no test owns it
+# (data providers, hook plumbing). Test-body stderr is merged into the captured
+# stdout and still surfaces in that test's own failure block.
+# Arguments: $1 - test file the worker ran, $2 - captured stderr file
+##
+function bashunit::console_results::print_worker_stderr() {
+ local test_file="$1"
+ local stderr_file="$2"
+
+ printf "\n%sStderr from %s%s\n" \
+ "$_BASHUNIT_COLOR_SKIPPED" "$test_file" "$_BASHUNIT_COLOR_DEFAULT"
+ sed 's/^/|/' "$stderr_file"
+}
+
 function bashunit::console_results::print_failing_tests_and_reset() {
  if [ -s "$FAILURES_OUTPUT_PATH" ]; then
  local total_failed

src/env.sh

@@ -654,6 +654,11 @@ 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"
+# Prefix for one file-per-worker capture of stderr written inside a parallel
+# worker but outside a test body. An ordinal is appended per spawned worker;
+# these must not live under TEMP_DIR_PARALLEL_TEST_SUITE, whose every entry is
+# walked by state::aggregate_parallel_results (#864).
+WORKER_STDERR_OUTPUT_PREFIX="$_BASHUNIT_RUN_OUTPUT_DIR/worker-stderr"
 # Collects "<test_file>:<function_name>" for every failing test in a run so the
 # next --rerun-failed can replay just those. Shared across parallel subshells.
 RERUN_FAILED_OUTPUT_PATH="$_BASHUNIT_RUN_OUTPUT_DIR/rerun-failed"

src/runner.sh

@@ -396,6 +396,9 @@ function bashunit::runner::load_test_files() {
  files=("$@")
  local -a scripts_ids=()
  local scripts_ids_count=0
+ local -a worker_stderr_paths=()
+ local -a worker_stderr_owners=()
+ local worker_stderr_count=0
 
  # Randomize file execution order (deterministic for the resolved seed).
  if bashunit::env::is_random_order_enabled; then
@@ -538,7 +541,15 @@ function bashunit::runner::load_test_files() {
  local _cached_fns="$functions_for_script"
  if bashunit::parallel::is_enabled; then
  bashunit::runner::wait_for_job_slot
- bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>/dev/null &
+ # Capture rather than discard: a worker's stderr cannot be written
+ # straight to the terminal without shredding the progress line, but
+ # dropping it made the same run report differently under --parallel
+ # (#358 added the discard, #864 replaced it with this capture).
+ local _worker_stderr="${WORKER_STDERR_OUTPUT_PREFIX}.${worker_stderr_count}"
+ worker_stderr_paths[worker_stderr_count]="$_worker_stderr"
+ worker_stderr_owners[worker_stderr_count]="$test_file"
+ worker_stderr_count=$((worker_stderr_count + 1))
+ bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>"$_worker_stderr" &
  else
  bashunit::runner::call_test_functions "$test_file" "$_cached_fns"
  fi
@@ -561,6 +572,16 @@ function bashunit::runner::load_test_files() {
  disown "$spinner_pid" 2>/dev/null || true
  kill "$spinner_pid" 2>/dev/null || true
  printf "\r \r" # Clear the spinner output
+
+ local _stderr_idx=0
+ while [ "$_stderr_idx" -lt "$worker_stderr_count" ]; do
+ if [ -s "${worker_stderr_paths[_stderr_idx]:-}" ]; then
+ bashunit::console_results::print_worker_stderr \
+ "${worker_stderr_owners[_stderr_idx]:-}" "${worker_stderr_paths[_stderr_idx]:-}"
+ fi
+ _stderr_idx=$((_stderr_idx + 1))
+ done
+
  local script_id
  for script_id in "${scripts_ids[@]+"${scripts_ids[@]}"}"; do
  export BASHUNIT_CURRENT_SCRIPT_ID="${script_id}"

tests/acceptance/fixtures/test_worker_stderr.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+
+# A data provider runs inside the parallel file worker but outside any test
+# body, so its stderr is the plain case that the worker-level redirect used to
+# discard. Test-body stderr is a different path: it is merged into the captured
+# stdout and already surfaces in the failure block.
+
+function data_provider_noisy() {
+ echo "WORKER-SCOPE-DIAGNOSTIC" >&2
+ echo "1"
+}
+
+# @data_provider data_provider_noisy
+function test_uses_a_noisy_provider() {
+ assert_equals "1" "$1"
+}

tests/acceptance/worker_stderr_test.sh

@@ -0,0 +1,42 @@
+#!/usr/bin/env bash
+
+# Parallel workers used to run with `2>/dev/null`, so anything written to
+# stderr inside the worker but outside a test body vanished, and the same run
+# reported differently depending on --parallel (#864).
+#
+# The fixture's provider is also executed by the test-counting pass in the main
+# shell, so merely *finding* the diagnostic proves nothing — it leaks in from
+# that pass either way. What the bug actually broke is the two modes agreeing,
+# so that is what these compare.
+
+FIXTURE="tests/acceptance/fixtures/test_worker_stderr.sh"
+
+function _count_diagnostic_in() {
+ local parallel_flag="$1"
+
+ ./bashunit "$parallel_flag" "$FIXTURE" 2>&1 |
+ grep -c "WORKER-SCOPE-DIAGNOSTIC"
+}
+
+function test_parallel_reports_worker_stderr_as_often_as_sequential() {
+ local sequential parallel
+ sequential="$(_count_diagnostic_in --no-parallel)"
+ parallel="$(_count_diagnostic_in --parallel)"
+
+ assert_not_equals "0" "$sequential"
+ assert_equals "$sequential" "$parallel"
+}
+
+function test_parallel_run_attributes_worker_stderr_to_its_file() {
+ local output
+ output=$(./bashunit --parallel "$FIXTURE" 2>&1)
+
+ assert_contains "Stderr from $FIXTURE" "$output"
+}
+
+function test_parallel_run_stays_quiet_when_no_worker_writes_stderr() {
+ local output
+ output=$(./bashunit --parallel tests/acceptance/fixtures/test_coverage_engine.sh 2>&1)
+
+ assert_not_contains "Stderr from" "$output"
+}

CHANGELOG.md

@@ -16,6 +16,7 @@
 - The `--parallel` unsupported-OS warning no longer claims Alpine is excluded: Alpine has been a supported parallel platform since the race conditions were fixed, the message was simply never updated
 
 ### Fixed
+- `--parallel` no longer discards stderr written inside a worker but outside a test body, which made the same run report differently depending on the mode. Each file worker had been spawned with `2>/dev/null` since #358 to keep its noise off the progress line; its stderr is now captured per file and rendered afterwards as a `Stderr from <file>` block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Output written by a test *body* was never affected — it is merged into that test's captured stdout and already appeared in its failure block (#864)
 - The minimum-bash gate now compares the minor version as well as the major, so it enforces whatever `BASHUNIT_MIN_BASH_VERSION` declares. It previously accepted any `3.x` regardless of the stated minimum, and a version string carrying a suffix (`5.2.37(1)-release`) is now parsed instead of tripping the comparison. The floor itself is unchanged at **Bash 3.0+**; `printf -v`, `+=` and `[[ =~ ]]` are now rejected in `src/` by the compatibility gate, since 3.0/3.1 lack the first two and 3.2 changed quoted-pattern semantics for the third
 - An empty entry in `.env` no longer overrides a value the caller exported or set on the command line. `.env` is sourced under `set -o allexport`, so every line was an unconditional assignment: merely *listing* a name blanked it, and `BASHUNIT_OUTPUT_FORMAT=tap ./bashunit` silently stopped working in any project whose `.env` mentioned that setting. An empty entry now means "not configured here"; an entry with a value still takes effect. This had been actively concealing defects — two of the bugs fixed in #879 were not reproducible from inside a repo checkout for exactly this reason (#865)
 - A malformed benchmark annotation is now an error instead of a silent fallback to the default: `@revs=abc` quietly ran a single revolution, `@its=abc` a single iteration, and `@max_ms=abc` dropped the threshold entirely so the benchmark could never fail — each reporting success while measuring something other than what was written (#884)

src/console_results.sh

@@ -559,6 +559,24 @@ function bashunit::console_results::print_error_test() {
  bashunit::state::print_line "error" "$line"
 }
 
+##
+# Render stderr a parallel worker wrote outside any test body.
+# A sequential run lets this straight through to the terminal; parallel workers
+# have it captured per file so concurrent writes cannot shred the progress
+# line. Attributed to the file, not a test: it is emitted where no test owns it
+# (data providers, hook plumbing). Test-body stderr is merged into the captured
+# stdout and still surfaces in that test's own failure block.
+# Arguments: $1 - test file the worker ran, $2 - captured stderr file
+##
+function bashunit::console_results::print_worker_stderr() {
+ local test_file="$1"
+ local stderr_file="$2"
+
+ printf "\n%sStderr from %s%s\n" \
+ "$_BASHUNIT_COLOR_SKIPPED" "$test_file" "$_BASHUNIT_COLOR_DEFAULT"
+ sed 's/^/|/' "$stderr_file"
+}
+
 function bashunit::console_results::print_failing_tests_and_reset() {
  if [ -s "$FAILURES_OUTPUT_PATH" ]; then
  local total_failed

src/env.sh

@@ -654,6 +654,11 @@ 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"
+# Prefix for one file-per-worker capture of stderr written inside a parallel
+# worker but outside a test body. An ordinal is appended per spawned worker;
+# these must not live under TEMP_DIR_PARALLEL_TEST_SUITE, whose every entry is
+# walked by state::aggregate_parallel_results (#864).
+WORKER_STDERR_OUTPUT_PREFIX="$_BASHUNIT_RUN_OUTPUT_DIR/worker-stderr"
 # Collects "<test_file>:<function_name>" for every failing test in a run so the
 # next --rerun-failed can replay just those. Shared across parallel subshells.
 RERUN_FAILED_OUTPUT_PATH="$_BASHUNIT_RUN_OUTPUT_DIR/rerun-failed"

src/runner.sh

@@ -396,6 +396,9 @@ function bashunit::runner::load_test_files() {
  files=("$@")
  local -a scripts_ids=()
  local scripts_ids_count=0
+ local -a worker_stderr_paths=()
+ local -a worker_stderr_owners=()
+ local worker_stderr_count=0
 
  # Randomize file execution order (deterministic for the resolved seed).
  if bashunit::env::is_random_order_enabled; then
@@ -538,7 +541,15 @@ function bashunit::runner::load_test_files() {
  local _cached_fns="$functions_for_script"
  if bashunit::parallel::is_enabled; then
  bashunit::runner::wait_for_job_slot
- bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>/dev/null &
+ # Capture rather than discard: a worker's stderr cannot be written
+ # straight to the terminal without shredding the progress line, but
+ # dropping it made the same run report differently under --parallel
+ # (#358 added the discard, #864 replaced it with this capture).
+ local _worker_stderr="${WORKER_STDERR_OUTPUT_PREFIX}.${worker_stderr_count}"
+ worker_stderr_paths[worker_stderr_count]="$_worker_stderr"
+ worker_stderr_owners[worker_stderr_count]="$test_file"
+ worker_stderr_count=$((worker_stderr_count + 1))
+ bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>"$_worker_stderr" &
  else
  bashunit::runner::call_test_functions "$test_file" "$_cached_fns"
  fi
@@ -561,6 +572,16 @@ function bashunit::runner::load_test_files() {
  disown "$spinner_pid" 2>/dev/null || true
  kill "$spinner_pid" 2>/dev/null || true
  printf "\r \r" # Clear the spinner output
+
+ local _stderr_idx=0
+ while [ "$_stderr_idx" -lt "$worker_stderr_count" ]; do
+ if [ -s "${worker_stderr_paths[_stderr_idx]:-}" ]; then
+ bashunit::console_results::print_worker_stderr \
+ "${worker_stderr_owners[_stderr_idx]:-}" "${worker_stderr_paths[_stderr_idx]:-}"
+ fi
+ _stderr_idx=$((_stderr_idx + 1))
+ done
+
  local script_id
  for script_id in "${scripts_ids[@]+"${scripts_ids[@]}"}"; do
  export BASHUNIT_CURRENT_SCRIPT_ID="${script_id}"

tests/acceptance/fixtures/test_worker_stderr.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+
+# A data provider runs inside the parallel file worker but outside any test
+# body, so its stderr is the plain case that the worker-level redirect used to
+# discard. Test-body stderr is a different path: it is merged into the captured
+# stdout and already surfaces in the failure block.
+
+function data_provider_noisy() {
+ echo "WORKER-SCOPE-DIAGNOSTIC" >&2
+ echo "1"
+}
+
+# @data_provider data_provider_noisy
+function test_uses_a_noisy_provider() {
+ assert_equals "1" "$1"
+}

tests/acceptance/worker_stderr_test.sh

@@ -0,0 +1,42 @@
+#!/usr/bin/env bash
+
+# Parallel workers used to run with `2>/dev/null`, so anything written to
+# stderr inside the worker but outside a test body vanished, and the same run
+# reported differently depending on --parallel (#864).
+#
+# The fixture's provider is also executed by the test-counting pass in the main
+# shell, so merely *finding* the diagnostic proves nothing — it leaks in from
+# that pass either way. What the bug actually broke is the two modes agreeing,
+# so that is what these compare.
+
+FIXTURE="tests/acceptance/fixtures/test_worker_stderr.sh"
+
+function _count_diagnostic_in() {
+ local parallel_flag="$1"
+
+ ./bashunit "$parallel_flag" "$FIXTURE" 2>&1 |
+ grep -c "WORKER-SCOPE-DIAGNOSTIC"
+}
+
+function test_parallel_reports_worker_stderr_as_often_as_sequential() {
+ local sequential parallel
+ sequential="$(_count_diagnostic_in --no-parallel)"
+ parallel="$(_count_diagnostic_in --parallel)"
+
+ assert_not_equals "0" "$sequential"
+ assert_equals "$sequential" "$parallel"
+}
+
+function test_parallel_run_attributes_worker_stderr_to_its_file() {
+ local output
+ output=$(./bashunit --parallel "$FIXTURE" 2>&1)
+
+ assert_contains "Stderr from $FIXTURE" "$output"
+}
+
+function test_parallel_run_stays_quiet_when_no_worker_writes_stderr() {
+ local output
+ output=$(./bashunit --parallel tests/acceptance/fixtures/test_coverage_engine.sh 2>&1)
+
+ assert_not_contains "Stderr from" "$output"
+}

CHANGELOG.md

@@ -16,6 +16,7 @@
 - The `--parallel` unsupported-OS warning no longer claims Alpine is excluded: Alpine has been a supported parallel platform since the race conditions were fixed, the message was simply never updated
 
 ### Fixed
+- `--parallel` no longer discards stderr written inside a worker but outside a test body, which made the same run report differently depending on the mode. Each file worker had been spawned with `2>/dev/null` since #358 to keep its noise off the progress line; its stderr is now captured per file and rendered afterwards as a `Stderr from <file>` block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Output written by a test *body* was never affected — it is merged into that test's captured stdout and already appeared in its failure block (#864)
 - The minimum-bash gate now compares the minor version as well as the major, so it enforces whatever `BASHUNIT_MIN_BASH_VERSION` declares. It previously accepted any `3.x` regardless of the stated minimum, and a version string carrying a suffix (`5.2.37(1)-release`) is now parsed instead of tripping the comparison. The floor itself is unchanged at **Bash 3.0+**; `printf -v`, `+=` and `[[ =~ ]]` are now rejected in `src/` by the compatibility gate, since 3.0/3.1 lack the first two and 3.2 changed quoted-pattern semantics for the third
 - An empty entry in `.env` no longer overrides a value the caller exported or set on the command line. `.env` is sourced under `set -o allexport`, so every line was an unconditional assignment: merely *listing* a name blanked it, and `BASHUNIT_OUTPUT_FORMAT=tap ./bashunit` silently stopped working in any project whose `.env` mentioned that setting. An empty entry now means "not configured here"; an entry with a value still takes effect. This had been actively concealing defects — two of the bugs fixed in #879 were not reproducible from inside a repo checkout for exactly this reason (#865)
 - A malformed benchmark annotation is now an error instead of a silent fallback to the default: `@revs=abc` quietly ran a single revolution, `@its=abc` a single iteration, and `@max_ms=abc` dropped the threshold entirely so the benchmark could never fail — each reporting success while measuring something other than what was written (#884)

src/console_results.sh

@@ -559,6 +559,24 @@ function bashunit::console_results::print_error_test() {
  bashunit::state::print_line "error" "$line"
 }
 
+##
+# Render stderr a parallel worker wrote outside any test body.
+# A sequential run lets this straight through to the terminal; parallel workers
+# have it captured per file so concurrent writes cannot shred the progress
+# line. Attributed to the file, not a test: it is emitted where no test owns it
+# (data providers, hook plumbing). Test-body stderr is merged into the captured
+# stdout and still surfaces in that test's own failure block.
+# Arguments: $1 - test file the worker ran, $2 - captured stderr file
+##
+function bashunit::console_results::print_worker_stderr() {
+ local test_file="$1"
+ local stderr_file="$2"
+
+ printf "\n%sStderr from %s%s\n" \
+ "$_BASHUNIT_COLOR_SKIPPED" "$test_file" "$_BASHUNIT_COLOR_DEFAULT"
+ sed 's/^/|/' "$stderr_file"
+}
+
 function bashunit::console_results::print_failing_tests_and_reset() {
  if [ -s "$FAILURES_OUTPUT_PATH" ]; then
  local total_failed

src/env.sh

@@ -654,6 +654,11 @@ 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"
+# Prefix for one file-per-worker capture of stderr written inside a parallel
+# worker but outside a test body. An ordinal is appended per spawned worker;
+# these must not live under TEMP_DIR_PARALLEL_TEST_SUITE, whose every entry is
+# walked by state::aggregate_parallel_results (#864).
+WORKER_STDERR_OUTPUT_PREFIX="$_BASHUNIT_RUN_OUTPUT_DIR/worker-stderr"
 # Collects "<test_file>:<function_name>" for every failing test in a run so the
 # next --rerun-failed can replay just those. Shared across parallel subshells.
 RERUN_FAILED_OUTPUT_PATH="$_BASHUNIT_RUN_OUTPUT_DIR/rerun-failed"

src/runner.sh

@@ -396,6 +396,9 @@ function bashunit::runner::load_test_files() {
  files=("$@")
  local -a scripts_ids=()
  local scripts_ids_count=0
+ local -a worker_stderr_paths=()
+ local -a worker_stderr_owners=()
+ local worker_stderr_count=0
 
  # Randomize file execution order (deterministic for the resolved seed).
  if bashunit::env::is_random_order_enabled; then
@@ -538,7 +541,15 @@ function bashunit::runner::load_test_files() {
  local _cached_fns="$functions_for_script"
  if bashunit::parallel::is_enabled; then
  bashunit::runner::wait_for_job_slot
- bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>/dev/null &
+ # Capture rather than discard: a worker's stderr cannot be written
+ # straight to the terminal without shredding the progress line, but
+ # dropping it made the same run report differently under --parallel
+ # (#358 added the discard, #864 replaced it with this capture).
+ local _worker_stderr="${WORKER_STDERR_OUTPUT_PREFIX}.${worker_stderr_count}"
+ worker_stderr_paths[worker_stderr_count]="$_worker_stderr"
+ worker_stderr_owners[worker_stderr_count]="$test_file"
+ worker_stderr_count=$((worker_stderr_count + 1))
+ bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>"$_worker_stderr" &
  else
  bashunit::runner::call_test_functions "$test_file" "$_cached_fns"
  fi
@@ -561,6 +572,16 @@ function bashunit::runner::load_test_files() {
  disown "$spinner_pid" 2>/dev/null || true
  kill "$spinner_pid" 2>/dev/null || true
  printf "\r \r" # Clear the spinner output
+
+ local _stderr_idx=0
+ while [ "$_stderr_idx" -lt "$worker_stderr_count" ]; do
+ if [ -s "${worker_stderr_paths[_stderr_idx]:-}" ]; then
+ bashunit::console_results::print_worker_stderr \
+ "${worker_stderr_owners[_stderr_idx]:-}" "${worker_stderr_paths[_stderr_idx]:-}"
+ fi
+ _stderr_idx=$((_stderr_idx + 1))
+ done
+
  local script_id
  for script_id in "${scripts_ids[@]+"${scripts_ids[@]}"}"; do
  export BASHUNIT_CURRENT_SCRIPT_ID="${script_id}"

tests/acceptance/fixtures/test_worker_stderr.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+
+# A data provider runs inside the parallel file worker but outside any test
+# body, so its stderr is the plain case that the worker-level redirect used to
+# discard. Test-body stderr is a different path: it is merged into the captured
+# stdout and already surfaces in the failure block.
+
+function data_provider_noisy() {
+ echo "WORKER-SCOPE-DIAGNOSTIC" >&2
+ echo "1"
+}
+
+# @data_provider data_provider_noisy
+function test_uses_a_noisy_provider() {
+ assert_equals "1" "$1"
+}

tests/acceptance/worker_stderr_test.sh

@@ -0,0 +1,42 @@
+#!/usr/bin/env bash
+
+# Parallel workers used to run with `2>/dev/null`, so anything written to
+# stderr inside the worker but outside a test body vanished, and the same run
+# reported differently depending on --parallel (#864).
+#
+# The fixture's provider is also executed by the test-counting pass in the main
+# shell, so merely *finding* the diagnostic proves nothing — it leaks in from
+# that pass either way. What the bug actually broke is the two modes agreeing,
+# so that is what these compare.
+
+FIXTURE="tests/acceptance/fixtures/test_worker_stderr.sh"
+
+function _count_diagnostic_in() {
+ local parallel_flag="$1"
+
+ ./bashunit "$parallel_flag" "$FIXTURE" 2>&1 |
+ grep -c "WORKER-SCOPE-DIAGNOSTIC"
+}
+
+function test_parallel_reports_worker_stderr_as_often_as_sequential() {
+ local sequential parallel
+ sequential="$(_count_diagnostic_in --no-parallel)"
+ parallel="$(_count_diagnostic_in --parallel)"
+
+ assert_not_equals "0" "$sequential"
+ assert_equals "$sequential" "$parallel"
+}
+
+function test_parallel_run_attributes_worker_stderr_to_its_file() {
+ local output
+ output=$(./bashunit --parallel "$FIXTURE" 2>&1)
+
+ assert_contains "Stderr from $FIXTURE" "$output"
+}
+
+function test_parallel_run_stays_quiet_when_no_worker_writes_stderr() {
+ local output
+ output=$(./bashunit --parallel tests/acceptance/fixtures/test_coverage_engine.sh 2>&1)
+
+ assert_not_contains "Stderr from" "$output"
+}

CHANGELOG.md

@@ -16,6 +16,7 @@
 - The `--parallel` unsupported-OS warning no longer claims Alpine is excluded: Alpine has been a supported parallel platform since the race conditions were fixed, the message was simply never updated
 
 ### Fixed
+- `--parallel` no longer discards stderr written inside a worker but outside a test body, which made the same run report differently depending on the mode. Each file worker had been spawned with `2>/dev/null` since #358 to keep its noise off the progress line; its stderr is now captured per file and rendered afterwards as a `Stderr from <file>` block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Output written by a test *body* was never affected — it is merged into that test's captured stdout and already appeared in its failure block (#864)
 - The minimum-bash gate now compares the minor version as well as the major, so it enforces whatever `BASHUNIT_MIN_BASH_VERSION` declares. It previously accepted any `3.x` regardless of the stated minimum, and a version string carrying a suffix (`5.2.37(1)-release`) is now parsed instead of tripping the comparison. The floor itself is unchanged at **Bash 3.0+**; `printf -v`, `+=` and `[[ =~ ]]` are now rejected in `src/` by the compatibility gate, since 3.0/3.1 lack the first two and 3.2 changed quoted-pattern semantics for the third
 - An empty entry in `.env` no longer overrides a value the caller exported or set on the command line. `.env` is sourced under `set -o allexport`, so every line was an unconditional assignment: merely *listing* a name blanked it, and `BASHUNIT_OUTPUT_FORMAT=tap ./bashunit` silently stopped working in any project whose `.env` mentioned that setting. An empty entry now means "not configured here"; an entry with a value still takes effect. This had been actively concealing defects — two of the bugs fixed in #879 were not reproducible from inside a repo checkout for exactly this reason (#865)
 - A malformed benchmark annotation is now an error instead of a silent fallback to the default: `@revs=abc` quietly ran a single revolution, `@its=abc` a single iteration, and `@max_ms=abc` dropped the threshold entirely so the benchmark could never fail — each reporting success while measuring something other than what was written (#884)

src/console_results.sh

@@ -559,6 +559,24 @@ function bashunit::console_results::print_error_test() {
  bashunit::state::print_line "error" "$line"
 }
 
+##
+# Render stderr a parallel worker wrote outside any test body.
+# A sequential run lets this straight through to the terminal; parallel workers
+# have it captured per file so concurrent writes cannot shred the progress
+# line. Attributed to the file, not a test: it is emitted where no test owns it
+# (data providers, hook plumbing). Test-body stderr is merged into the captured
+# stdout and still surfaces in that test's own failure block.
+# Arguments: $1 - test file the worker ran, $2 - captured stderr file
+##
+function bashunit::console_results::print_worker_stderr() {
+ local test_file="$1"
+ local stderr_file="$2"
+
+ printf "\n%sStderr from %s%s\n" \
+ "$_BASHUNIT_COLOR_SKIPPED" "$test_file" "$_BASHUNIT_COLOR_DEFAULT"
+ sed 's/^/|/' "$stderr_file"
+}
+
 function bashunit::console_results::print_failing_tests_and_reset() {
  if [ -s "$FAILURES_OUTPUT_PATH" ]; then
  local total_failed

src/env.sh

@@ -654,6 +654,11 @@ 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"
+# Prefix for one file-per-worker capture of stderr written inside a parallel
+# worker but outside a test body. An ordinal is appended per spawned worker;
+# these must not live under TEMP_DIR_PARALLEL_TEST_SUITE, whose every entry is
+# walked by state::aggregate_parallel_results (#864).
+WORKER_STDERR_OUTPUT_PREFIX="$_BASHUNIT_RUN_OUTPUT_DIR/worker-stderr"
 # Collects "<test_file>:<function_name>" for every failing test in a run so the
 # next --rerun-failed can replay just those. Shared across parallel subshells.
 RERUN_FAILED_OUTPUT_PATH="$_BASHUNIT_RUN_OUTPUT_DIR/rerun-failed"

src/runner.sh

@@ -396,6 +396,9 @@ function bashunit::runner::load_test_files() {
  files=("$@")
  local -a scripts_ids=()
  local scripts_ids_count=0
+ local -a worker_stderr_paths=()
+ local -a worker_stderr_owners=()
+ local worker_stderr_count=0
 
  # Randomize file execution order (deterministic for the resolved seed).
  if bashunit::env::is_random_order_enabled; then
@@ -538,7 +541,15 @@ function bashunit::runner::load_test_files() {
  local _cached_fns="$functions_for_script"
  if bashunit::parallel::is_enabled; then
  bashunit::runner::wait_for_job_slot
- bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>/dev/null &
+ # Capture rather than discard: a worker's stderr cannot be written
+ # straight to the terminal without shredding the progress line, but
+ # dropping it made the same run report differently under --parallel
+ # (#358 added the discard, #864 replaced it with this capture).
+ local _worker_stderr="${WORKER_STDERR_OUTPUT_PREFIX}.${worker_stderr_count}"
+ worker_stderr_paths[worker_stderr_count]="$_worker_stderr"
+ worker_stderr_owners[worker_stderr_count]="$test_file"
+ worker_stderr_count=$((worker_stderr_count + 1))
+ bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>"$_worker_stderr" &
  else
  bashunit::runner::call_test_functions "$test_file" "$_cached_fns"
  fi
@@ -561,6 +572,16 @@ function bashunit::runner::load_test_files() {
  disown "$spinner_pid" 2>/dev/null || true
  kill "$spinner_pid" 2>/dev/null || true
  printf "\r \r" # Clear the spinner output
+
+ local _stderr_idx=0
+ while [ "$_stderr_idx" -lt "$worker_stderr_count" ]; do
+ if [ -s "${worker_stderr_paths[_stderr_idx]:-}" ]; then
+ bashunit::console_results::print_worker_stderr \
+ "${worker_stderr_owners[_stderr_idx]:-}" "${worker_stderr_paths[_stderr_idx]:-}"
+ fi
+ _stderr_idx=$((_stderr_idx + 1))
+ done
+
  local script_id
  for script_id in "${scripts_ids[@]+"${scripts_ids[@]}"}"; do
  export BASHUNIT_CURRENT_SCRIPT_ID="${script_id}"

tests/acceptance/fixtures/test_worker_stderr.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+
+# A data provider runs inside the parallel file worker but outside any test
+# body, so its stderr is the plain case that the worker-level redirect used to
+# discard. Test-body stderr is a different path: it is merged into the captured
+# stdout and already surfaces in the failure block.
+
+function data_provider_noisy() {
+ echo "WORKER-SCOPE-DIAGNOSTIC" >&2
+ echo "1"
+}
+
+# @data_provider data_provider_noisy
+function test_uses_a_noisy_provider() {
+ assert_equals "1" "$1"
+}

tests/acceptance/worker_stderr_test.sh

@@ -0,0 +1,42 @@
+#!/usr/bin/env bash
+
+# Parallel workers used to run with `2>/dev/null`, so anything written to
+# stderr inside the worker but outside a test body vanished, and the same run
+# reported differently depending on --parallel (#864).
+#
+# The fixture's provider is also executed by the test-counting pass in the main
+# shell, so merely *finding* the diagnostic proves nothing — it leaks in from
+# that pass either way. What the bug actually broke is the two modes agreeing,
+# so that is what these compare.
+
+FIXTURE="tests/acceptance/fixtures/test_worker_stderr.sh"
+
+function _count_diagnostic_in() {
+ local parallel_flag="$1"
+
+ ./bashunit "$parallel_flag" "$FIXTURE" 2>&1 |
+ grep -c "WORKER-SCOPE-DIAGNOSTIC"
+}
+
+function test_parallel_reports_worker_stderr_as_often_as_sequential() {
+ local sequential parallel
+ sequential="$(_count_diagnostic_in --no-parallel)"
+ parallel="$(_count_diagnostic_in --parallel)"
+
+ assert_not_equals "0" "$sequential"
+ assert_equals "$sequential" "$parallel"
+}
+
+function test_parallel_run_attributes_worker_stderr_to_its_file() {
+ local output
+ output=$(./bashunit --parallel "$FIXTURE" 2>&1)
+
+ assert_contains "Stderr from $FIXTURE" "$output"
+}
+
+function test_parallel_run_stays_quiet_when_no_worker_writes_stderr() {
+ local output
+ output=$(./bashunit --parallel tests/acceptance/fixtures/test_coverage_engine.sh 2>&1)
+
+ assert_not_contains "Stderr from" "$output"
+}

CHANGELOG.md

@@ -16,6 +16,7 @@
 - The `--parallel` unsupported-OS warning no longer claims Alpine is excluded: Alpine has been a supported parallel platform since the race conditions were fixed, the message was simply never updated
 
 ### Fixed
+- `--parallel` no longer discards stderr written inside a worker but outside a test body, which made the same run report differently depending on the mode. Each file worker had been spawned with `2>/dev/null` since #358 to keep its noise off the progress line; its stderr is now captured per file and rendered afterwards as a `Stderr from <file>` block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Output written by a test *body* was never affected — it is merged into that test's captured stdout and already appeared in its failure block (#864)
 - The minimum-bash gate now compares the minor version as well as the major, so it enforces whatever `BASHUNIT_MIN_BASH_VERSION` declares. It previously accepted any `3.x` regardless of the stated minimum, and a version string carrying a suffix (`5.2.37(1)-release`) is now parsed instead of tripping the comparison. The floor itself is unchanged at **Bash 3.0+**; `printf -v`, `+=` and `[[ =~ ]]` are now rejected in `src/` by the compatibility gate, since 3.0/3.1 lack the first two and 3.2 changed quoted-pattern semantics for the third
 - An empty entry in `.env` no longer overrides a value the caller exported or set on the command line. `.env` is sourced under `set -o allexport`, so every line was an unconditional assignment: merely *listing* a name blanked it, and `BASHUNIT_OUTPUT_FORMAT=tap ./bashunit` silently stopped working in any project whose `.env` mentioned that setting. An empty entry now means "not configured here"; an entry with a value still takes effect. This had been actively concealing defects — two of the bugs fixed in #879 were not reproducible from inside a repo checkout for exactly this reason (#865)
 - A malformed benchmark annotation is now an error instead of a silent fallback to the default: `@revs=abc` quietly ran a single revolution, `@its=abc` a single iteration, and `@max_ms=abc` dropped the threshold entirely so the benchmark could never fail — each reporting success while measuring something other than what was written (#884)

src/console_results.sh

@@ -559,6 +559,24 @@ function bashunit::console_results::print_error_test() {
  bashunit::state::print_line "error" "$line"
 }
 
+##
+# Render stderr a parallel worker wrote outside any test body.
+# A sequential run lets this straight through to the terminal; parallel workers
+# have it captured per file so concurrent writes cannot shred the progress
+# line. Attributed to the file, not a test: it is emitted where no test owns it
+# (data providers, hook plumbing). Test-body stderr is merged into the captured
+# stdout and still surfaces in that test's own failure block.
+# Arguments: $1 - test file the worker ran, $2 - captured stderr file
+##
+function bashunit::console_results::print_worker_stderr() {
+ local test_file="$1"
+ local stderr_file="$2"
+
+ printf "\n%sStderr from %s%s\n" \
+ "$_BASHUNIT_COLOR_SKIPPED" "$test_file" "$_BASHUNIT_COLOR_DEFAULT"
+ sed 's/^/|/' "$stderr_file"
+}
+
 function bashunit::console_results::print_failing_tests_and_reset() {
  if [ -s "$FAILURES_OUTPUT_PATH" ]; then
  local total_failed

src/env.sh

@@ -654,6 +654,11 @@ 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"
+# Prefix for one file-per-worker capture of stderr written inside a parallel
+# worker but outside a test body. An ordinal is appended per spawned worker;
+# these must not live under TEMP_DIR_PARALLEL_TEST_SUITE, whose every entry is
+# walked by state::aggregate_parallel_results (#864).
+WORKER_STDERR_OUTPUT_PREFIX="$_BASHUNIT_RUN_OUTPUT_DIR/worker-stderr"
 # Collects "<test_file>:<function_name>" for every failing test in a run so the
 # next --rerun-failed can replay just those. Shared across parallel subshells.
 RERUN_FAILED_OUTPUT_PATH="$_BASHUNIT_RUN_OUTPUT_DIR/rerun-failed"

src/runner.sh

@@ -396,6 +396,9 @@ function bashunit::runner::load_test_files() {
  files=("$@")
  local -a scripts_ids=()
  local scripts_ids_count=0
+ local -a worker_stderr_paths=()
+ local -a worker_stderr_owners=()
+ local worker_stderr_count=0
 
  # Randomize file execution order (deterministic for the resolved seed).
  if bashunit::env::is_random_order_enabled; then
@@ -538,7 +541,15 @@ function bashunit::runner::load_test_files() {
  local _cached_fns="$functions_for_script"
  if bashunit::parallel::is_enabled; then
  bashunit::runner::wait_for_job_slot
- bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>/dev/null &
+ # Capture rather than discard: a worker's stderr cannot be written
+ # straight to the terminal without shredding the progress line, but
+ # dropping it made the same run report differently under --parallel
+ # (#358 added the discard, #864 replaced it with this capture).
+ local _worker_stderr="${WORKER_STDERR_OUTPUT_PREFIX}.${worker_stderr_count}"
+ worker_stderr_paths[worker_stderr_count]="$_worker_stderr"
+ worker_stderr_owners[worker_stderr_count]="$test_file"
+ worker_stderr_count=$((worker_stderr_count + 1))
+ bashunit::runner::call_test_functions "$test_file" "$_cached_fns" 2>"$_worker_stderr" &
  else
  bashunit::runner::call_test_functions "$test_file" "$_cached_fns"
  fi
@@ -561,6 +572,16 @@ function bashunit::runner::load_test_files() {
  disown "$spinner_pid" 2>/dev/null || true
  kill "$spinner_pid" 2>/dev/null || true
  printf "\r \r" # Clear the spinner output
+
+ local _stderr_idx=0
+ while [ "$_stderr_idx" -lt "$worker_stderr_count" ]; do
+ if [ -s "${worker_stderr_paths[_stderr_idx]:-}" ]; then
+ bashunit::console_results::print_worker_stderr \
+ "${worker_stderr_owners[_stderr_idx]:-}" "${worker_stderr_paths[_stderr_idx]:-}"
+ fi
+ _stderr_idx=$((_stderr_idx + 1))
+ done
+
  local script_id
  for script_id in "${scripts_ids[@]+"${scripts_ids[@]}"}"; do
  export BASHUNIT_CURRENT_SCRIPT_ID="${script_id}"

tests/acceptance/fixtures/test_worker_stderr.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+
+# A data provider runs inside the parallel file worker but outside any test
+# body, so its stderr is the plain case that the worker-level redirect used to
+# discard. Test-body stderr is a different path: it is merged into the captured
+# stdout and already surfaces in the failure block.
+
+function data_provider_noisy() {
+ echo "WORKER-SCOPE-DIAGNOSTIC" >&2
+ echo "1"
+}
+
+# @data_provider data_provider_noisy
+function test_uses_a_noisy_provider() {
+ assert_equals "1" "$1"
+}

tests/acceptance/worker_stderr_test.sh

@@ -0,0 +1,42 @@
+#!/usr/bin/env bash
+
+# Parallel workers used to run with `2>/dev/null`, so anything written to
+# stderr inside the worker but outside a test body vanished, and the same run
+# reported differently depending on --parallel (#864).
+#
+# The fixture's provider is also executed by the test-counting pass in the main
+# shell, so merely *finding* the diagnostic proves nothing — it leaks in from
+# that pass either way. What the bug actually broke is the two modes agreeing,
+# so that is what these compare.
+
+FIXTURE="tests/acceptance/fixtures/test_worker_stderr.sh"
+
+function _count_diagnostic_in() {
+ local parallel_flag="$1"
+
+ ./bashunit "$parallel_flag" "$FIXTURE" 2>&1 |
+ grep -c "WORKER-SCOPE-DIAGNOSTIC"
+}
+
+function test_parallel_reports_worker_stderr_as_often_as_sequential() {
+ local sequential parallel
+ sequential="$(_count_diagnostic_in --no-parallel)"
+ parallel="$(_count_diagnostic_in --parallel)"
+
+ assert_not_equals "0" "$sequential"
+ assert_equals "$sequential" "$parallel"
+}
+
+function test_parallel_run_attributes_worker_stderr_to_its_file() {
+ local output
+ output=$(./bashunit --parallel "$FIXTURE" 2>&1)
+
+ assert_contains "Stderr from $FIXTURE" "$output"
+}
+
+function test_parallel_run_stays_quiet_when_no_worker_writes_stderr() {
+ local output
+ output=$(./bashunit --parallel tests/acceptance/fixtures/test_coverage_engine.sh 2>&1)
+
+ assert_not_contains "Stderr from" "$output"
+}