perf: reduce acceptance-suite cold-start cost (258 nested ./bashunit runs)

[Chemaclass]

perf: reduce acceptance-suite cold-start cost (258 nested ./bashunit runs)

Follow-up to #761 (rounds 1+2 complete). The per-test hot path is now largely fork-free; the dominant remaining wall-clock cost is elsewhere.

Problem

The acceptance suite spawns 258 nested ./bashunit subprocess runs. Each pays a full cold start (sourcing all of src/ + env resolution). At ~52s for the full --parallel suite, these cold starts — not per-test overhead — dominate.

Candidate directions (each needs profiling first)

• Trim cold-start cost — sourcing + env resolution per invocation. Sourcing all of src/ measures only ~28ms, so this alone caps at ~5-8s across 258 runs.
• Share one built artifact across nested runs instead of re-sourcing each time.
• Reclassify tests — convert acceptance tests that don't exercise the CLI surface into functional tests (no nested process needed).

Measure before filing sub-work

Profile a representative nested run to attribute cold-start cost (source vs env vs fork) before committing to a direction. Report numbers, then pick.

Also considered, NOT pursued (from #761)

Lazy-sourcing large optional modules (coverage.sh, learn.sh): sourcing all of src/ is only ~28ms, so the win is ~5-8s across 258 runs and it complicates the single-file release build. Revisit only inside the cold-start work above.

Constraints

• Bash 3.0+ compatible (target machine is bash 3.2 on macOS): no declare -A, [[ ]], ${var,,}, negative array indexing, &>>.
• TDD; quality gate (./bashunit tests/, ./bashunit --parallel tests/, make sa, make lint, shfmt -w .) on every PR.
• One issue per PR.

Benchmark protocol

mkdir -p /tmp/bashunit-bench
for i in $(seq 1 100); do echo "function test_noop_$i() { assert_true true; }"; done > /tmp/bashunit-bench/noop100_test.sh
time ./bashunit --no-parallel /tmp/bashunit-bench/noop100_test.sh

# fork census
PS4='+X ${BASH_SOURCE##*/}:$LINENO ' bash -x ./bashunit --no-parallel /tmp/bashunit-bench/noop100_test.sh \
  2>/tmp/bashunit-bench/trace.txt >/dev/null
grep -oE ' (perl|base64|grep|sed|mktemp|rm|date) ' /tmp/bashunit-bench/trace.txt | sort | uniq -c

[Chemaclass]

Cold-start profiling (bash 3.2.57, Apple Silicon)

Ran the measure-first pass this issue calls for. Single trivial nested run:

• Cold start: ~118ms self-reported per ./bashunit --no-parallel <one test>.
• Sourcing all of src/: ~25ms — confirms the earlier ~28ms figure; sourcing is a small slice of the total.
• Startup external forks (per cold start, before any test body runs): a handful — perl (suite clock, unavoidable on bash 3.2), uname, base64, dirname, sort. Long-tail, not one dominant offender.

Landed

• uname was forked twice per load (check_os.sh module scope + check_os::init). Deduped to one in perf(runner): single uname fork at cold start #799. Small but on every one of the 258 nested runs.

Still open (the real levers, unchanged)

• Share one built artifact across nested runs instead of re-sourcing — largest potential win, needs a design.
• Reclassify acceptance tests that never exercise the CLI surface into functional tests (removes the nested process entirely).
• Further cold-start fork trimming is long-tail; each fork saves ~1-2ms x 258 runs.

Keeping this open as the frontier tracker; #799 is a partial win, not a close.