`resolve_test_location` forks a subshell per test for a line number the per-file scan already read (1.08ms per test)

[Chemaclass]

Summary

Every test forks a subshell to read a line number, and the per-file scan already read it.

Measured

$(shopt -s extdebug; declare -F fn) costs 1.08 ms per call on macOS arm64, bash 3.2.57, best of 3 over 2000 calls. It runs once per test, unconditionally. Over this suite's ~2491 tests that is roughly 2.7 seconds of a sequential run.

A bare $( ) costs 1.1 to 1.4 ms on the same box. Worth noting: perf-fork-budget.md records 0.46 ms, which predates arm64. The old figure should not be reused for estimates.

Current behavior

bashunit::runner::resolve_test_location, src/runner/context.sh:80-100:

def="$(shopt -s extdebug; declare -F "$fn_name" 2>/dev/null)" || true

It is called from export_test_identity (src/runner/context.sh:65) for every test, to build <file>:<line> for a failure message most tests never emit.

Meanwhile bashunit::runner::functions_for_script (src/runner/discovery.sh:340-380) already runs $( shopt -s extdebug; declare -F $all_fn_names ) once per file, reads <name> <line> <file> for every function in it, sorts by definition line, and then throws the line numbers away.

Expected behavior

Cache name to line in that existing per-file pass and make the per-test lookup a scan of it. Bash 3 parallel arrays, following _BASHUNIT_TAGS_MAP_FNS / _BASHUNIT_TAGS_MAP_TAGS in src/helper/tags.sh. Keep the current $( ) as the fallback for a name that is not in the map, so nothing regresses for a function defined some other way.

extdebug must stay inside the subshell (#808). Verify against a data-provider test, where one function runs for many rows, and against an interpolated title.

[Chemaclass]

Hazard for any rewrite here: extdebug clears set -E and set -T on Bash 5.2+

Worth knowing before touching this path. shopt -s extdebug turns on errtrace and functrace everywhere, but turning it off is not symmetric:

Build	after shopt -u extdebug
3.00.22, 3.2.57, 4.0.44	errtrace/functrace stay on
5.2.37, 5.3.15	both off, even if they were on beforehand

So on 5.2 and above, disabling extdebug silently turns off set -E and set -T. That is the concrete shape of the #808 hazard, and it differs across the supported range, which the comment at src/runner/context.sh:84-88 does not say.

Consequence for this issue: keeping extdebug inside a subshell is not only about not clobbering caller state, it is also what stops --strict error tracing being switched off underneath the run on newer shells. A rewrite that drops the subshell has to save and restore errtrace/functrace explicitly. Tracked in #1354.

Measured saving, for scale

Removing this subshell measured 0.27 to 0.39 ms/test on Linux and 0.62 ms on macOS. Doing it the Bash 3.0-safe way beat rewriting it with Bash 5.3's ${ cmd; }: the gate-free pair of this plus #1348 saved 0.64 ms/test on 5.3, where the fully gated version of the same sites saved 0.60 ms. So no version gate is wanted here.

[Chemaclass]

Done in #1356 (merged).