perf(runner): ordinal result-file naming in parallel workers (drop mktemp+mv per test)

[Chemaclass]

🎯 Goal

Eliminate the per-test mktemp + mv (and, for data-provider tests, the extra echo | tr | sed sanitize pipeline) that every --parallel worker forks when writing its .result file. Replace the unique-filename dance with a per-suite monotonic ordinal minted in the (single-threaded) dispatch loop.

This is the single largest remaining chunk of the parallel fork budget: ~10 mktemp + 10 mv on a 10-test file today, plus up to 3 more forks per provider-argument test.

📍 Where

• Result file is written at the tail of bashunit::runner::run_test — src/runner.sh:1595-1622:
  • template="${fn_name}.XXXXXX" (or ${fn_name}-${sanitized_args}.XXXXXX)
  • unique_test_result_file=$("$MKTEMP" -p "$test_suite_dir" "$template") → fork
  • mv "$unique_test_result_file" "${unique_test_result_file}.result" → fork
  • sanitized_args=$(echo "${args[*]}" | tr … | sed …) → 3 forks, provider tests only (src/runner.sh:1600)
• Callers loop sequentially inside one worker (one file = one worker): bashunit::runner::call_test_functions at src/runner.sh:867 (fn loop) and :896 (provider-data loop), calling run_test at :877/:879/:907/:909.
• Result files live in a per-suite subdir: ${TEMP_DIR_PARALLEL_TEST_SUITE}/${test_suite_base%.sh} (src/runner.sh:864,1595).
• Aggregation globs *.result in that dir: bashunit::parallel::aggregate_test_results (src/runner.sh:522).

💡 Why it is safe (design)

Parallelism is per file, not per test — call_test_functions is what gets backgrounded with & (src/runner.sh:499-502). Inside a worker, tests run sequentially, and every worker writes into its own per-suite subdir. Therefore a counter that increments once per run_test call within a suite is unique within that subdir with no cross-worker collision — which is exactly why deterministic names were previously rejected (they were keyed on the sanitized test name, and different provider args can sanitize identically). An ordinal sidesteps that.

🔧 Approach

1. In call_test_functions, keep a local _test_index=0; increment before each run_test call (both the plain loop at :867 and the provider loop at :896).
2. Pass the index to run_test (new trailing/leading parameter — keep the existing positional args intact; prefer a dedicated global slot or an explicit param, not reuse of $@ which carries provider data).
3. In run_test's parallel tail, replace the whole template/mktemp/mv/sanitized_args block with:

   local unique_test_result_file="${test_suite_dir}/${_test_index}.result"

   No mktemp, no mv, no sanitize pipeline. echo "$execution_result" >"$unique_test_result_file" stays.
4. Confirm nothing else parses the old fn_name/sanitized_args filename shape (aggregation only globs *.result; verify src/parallel.sh).

✅ Acceptance criteria

• --parallel runs produce one <N>.result per test in the per-suite subdir; aggregation totals are byte-identical to before.
• Zero mktemp/mv forks on the per-test parallel path; zero echo|tr|sed for provider tests.
• Sequential path unchanged.

🧪 TDD / measurement

• RED first: extend the parallel budget assertion in tests/acceptance/bashunit_run_forks_test.sh to pin the new (lower) mktemp/mv counts — it must fail before the change. Use the PATH-shim census method in .claude/rules/perf-fork-budget.md (note: $MKTEMP is pinned at startup and bypasses PATH shims — shim mktemp/mv accordingly or trace).
• Add/adjust a functional test asserting a 10-test file + a data-provider file still aggregate correct pass/fail counts under --parallel.
• Full gate: ./bashunit tests/ and ./bashunit --parallel --simple --strict tests/, make sa, make lint.

⛓️ Constraints

• Bash 3.0+ (.claude/rules/bash-style.md): no [[, declare -A, ${var,,}, negative indices, BASHPID. A monotonic local counter is fine — it lives in the single worker shell, no shared-$$ problem.
• Follow the return-slot pattern if threading the index via a global (_BASHUNIT_*).
• Update .claude/rules/perf-fork-budget.md "Current budgets" table to reflect the new parallel count.

📈 Impact

Roughly halves the per-10-test-file parallel fork count; larger win on data-provider-heavy suites.

[Chemaclass]

Done in #857 (merged).