perf(runner): build the provider map once per file, not twice

The data-provider map was scanned by awk twice per test file: once inside the
header's test-counting subshell (whose single-entry cache died with it) and
again by the runner. The count loop now builds the map in its own shell before
entering the counting subshell, and the header reads the total through a
return slot instead of a $(...) capture — so the cache survives into the
runner, which hits it for single-file runs (the acceptance suite's ~258 nested
runs are all single-file). Multi-file runs keep today's behaviour for all but
the last file (single-entry cache).

3 -> 2 awk forks per single-file run; stdout contract of find_total_tests is
unchanged for existing callers/tests.

Author: Chemaclass

CHANGELOG.md

@@ -15,6 +15,7 @@
 - `--jobs auto` / `-j auto` caps parallel concurrency at the CPU core count (portable across Linux/macOS/BSD); the default stays unlimited (#766)
 
 ### Changed
+- Faster test runs: the data-provider map is built once per file in the main shell (the header's test count reads a return slot instead of a `$(...)` capture, so the cache survives for the runner) — one awk scan per file instead of two
 - Faster test runs: stripping ANSI codes from short colored strings is pure bash now, so aligning the per-test execution time (shown on systems with a fork-free clock, e.g. Linux) no longer forks `sed` once per passing test
 - Faster parallel runs: publishing each test's result file no longer forks `basename`, `mkdir` and an `echo | tr | sed` pipeline per test — the suite dir name is parameter expansion, the dir is pre-created once per file before workers spawn, and arg sanitizing is skipped without provider args (a 10-test parallel run dropped from 61 to 21 forks). No behaviour change
 - Faster test runs: listing all defined functions uses the `compgen -A function` builtin instead of forking `declare -F | awk` (three call sites; 5 -> 3 `awk` forks per test file). No behaviour change

src/console_header.sh

@@ -39,7 +39,11 @@ function bashunit::console_header::print_version() {
     # Skip counting in parallel+simple mode for faster startup
     total_tests=0
   else
-    total_tests=$(bashunit::helper::find_total_tests "$filter" "$@")
+    # Read via the return slot, not $(...): the capture subshell would discard
+    # the provider-map cache find_total_tests builds per file, forcing the
+    # runner to re-scan each file with a second awk fork.
+    bashunit::helper::find_total_tests "$filter" "$@" >/dev/null
+    total_tests=$_BASHUNIT_HELPER_TOTAL_TESTS_OUT
   fi
 
   if bashunit::env::is_header_ascii_art_enabled; then

src/helpers.sh

@@ -526,10 +526,15 @@ function bashunit::helper::get_latest_tag() {
     head -n 1
 }
 
+# Also written by find_total_tests so a main-shell caller can read the count
+# without a $() capture (which would discard the provider-map cache built here).
+_BASHUNIT_HELPER_TOTAL_TESTS_OUT=0
+
 function bashunit::helper::find_total_tests() {
   local filter=${1:-}
   shift || true
 
+  _BASHUNIT_HELPER_TOTAL_TESTS_OUT=0
   if [ $# -eq 0 ]; then
     echo 0
     return
@@ -543,6 +548,12 @@ function bashunit::helper::find_total_tests() {
       continue
     fi
 
+    # Build the provider map in THIS shell before the counting subshell: the
+    # subshell inherits it (its own build call becomes a cache hit), and when
+    # the caller runs in the main shell the runner's later build for the same
+    # file is a cache hit too — one awk scan per file instead of two.
+    bashunit::helper::build_provider_map "$file"
+
     local file_count
     file_count=$( (
       # shellcheck source=/dev/null
@@ -590,6 +601,7 @@ function bashunit::helper::find_total_tests() {
     total_count=$((total_count + file_count))
   done
 
+  _BASHUNIT_HELPER_TOTAL_TESTS_OUT=$total_count
   echo "$total_count"
 }
 

tests/acceptance/bashunit_run_forks_test.sh

@@ -126,8 +126,9 @@ function test_running_a_test_file_does_not_fork_sort() {
 
 # Regression guard: listing all defined functions must use the `compgen -A
 # function` builtin, not a `declare -F | awk` fork. The remaining awk budget of
-# a plain run is the per-file file scans (data-provider map, which runs in both
-# the counting subshell and the runner, plus the duplicate-name check).
+# a single-file run is one data-provider scan (built once in the main shell so
+# the header-count subshell and the runner both hit the cache) plus the
+# duplicate-name check.
 function test_running_a_test_file_stays_within_the_awk_fork_budget() {
   if bashunit::check_os::is_windows; then
     bashunit::skip "PATH shims are unreliable under Git Bash" && return
@@ -155,7 +156,7 @@ function test_running_a_test_file_stays_within_the_awk_fork_budget() {
     awk_forks="$(grep -c . "$count_file" || true)"
   fi
 
-  assert_less_or_equal_than 3 "$awk_forks"
+  assert_less_or_equal_than 2 "$awk_forks"
 }
 
 # Regression guard: a run must remove its run-output scratch directory on exit.