fix(parallel): key per-test result dirs on the path, not the basename

[Chemaclass]

🤔 Background

Closes#959

Under --parallel, each test file's results were bucketed into a directory named by the file's basename:

local test_suite_base="${test_file##*/}"local test_suite_dir="${TEMP_DIR_PARALLEL_TEST_SUITE}/${test_suite_base%.sh}"

Two files sharing a basename in different directories shared one bucket. The per-suite ordinal is unique per suite, not globally, so <tmp>/foo_test/1.result was written twice and the first write was lost. The run stayed green and simply reported fewer tests.

Same defect as #923, which fixed exactly this in build.sh: a basename isn't an identity when directories can repeat one.

📉 How it was found, and why it matters to users

Mirroring tests/unit/ onto the src/ module layout (#957/#960) produced three duplicate basenames and cost 52 tests under --parallel:

	--parallel --simple --strict
with 3 duplicate basenames	1536
after renaming them unique	1588

A tests/ tree mirroring a src/ tree is what the docs encourage, so any user project doing it is exposed — and a failing test can vanish while CI stays green.

💡 Changes

• The derivation folds the whole path instead of taking the basename.
• Both sites that computed it — runner/result.sh and the pre-create in runner/exec.sh (perf(runner): stop forking per test when publishing parallel results #813) — now call one shared helper, so they cannot drift apart.
• It stays pure parameter expansion: this runs once per test in every parallel worker and the path must remain fork-free. The fork-budget acceptance tests pass unchanged (11/11).

Nothing parses the directory name — aggregate_parallel_results globs whatever subdirectories exist — so the rename is invisible downstream.

✅ Verification

TDD: the new acceptance test in bashunit_parallel_consistency_test.sh runs two same-named fixture files from different directories in both modes and compares counts. Verified red without the fix (sequential 2, parallel 1) and green with it.

It asserts on extracted counts rather than the rendered summary — the parallel spinner leaves control bytes on that line — plus a guard so both sides can't collapse to empty and match vacuously.

Green: sequential (1630) · --parallel --simple --strict (1589) · fork budget · make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅. CHANGELOG updated under Fixed.

[Chemaclass]

🤔 Background

Closes#959

Under --parallel, each test file's results were bucketed into a directory named by the file's basename:

local test_suite_base="${test_file##*/}"local test_suite_dir="${TEMP_DIR_PARALLEL_TEST_SUITE}/${test_suite_base%.sh}"

Two files sharing a basename in different directories shared one bucket. The per-suite ordinal is unique per suite, not globally, so <tmp>/foo_test/1.result was written twice and the first write was lost. The run stayed green and simply reported fewer tests.

Same defect as #923, which fixed exactly this in build.sh: a basename isn't an identity when directories can repeat one.

📉 How it was found, and why it matters to users

Mirroring tests/unit/ onto the src/ module layout (#957/#960) produced three duplicate basenames and cost 52 tests under --parallel:

	--parallel --simple --strict
with 3 duplicate basenames	1536
after renaming them unique	1588

A tests/ tree mirroring a src/ tree is what the docs encourage, so any user project doing it is exposed — and a failing test can vanish while CI stays green.

💡 Changes

• The derivation folds the whole path instead of taking the basename.
• Both sites that computed it — runner/result.sh and the pre-create in runner/exec.sh (perf(runner): stop forking per test when publishing parallel results #813) — now call one shared helper, so they cannot drift apart.
• It stays pure parameter expansion: this runs once per test in every parallel worker and the path must remain fork-free. The fork-budget acceptance tests pass unchanged (11/11).

Nothing parses the directory name — aggregate_parallel_results globs whatever subdirectories exist — so the rename is invisible downstream.

✅ Verification

TDD: the new acceptance test in bashunit_parallel_consistency_test.sh runs two same-named fixture files from different directories in both modes and compares counts. Verified red without the fix (sequential 2, parallel 1) and green with it.

It asserts on extracted counts rather than the rendered summary — the parallel spinner leaves control bytes on that line — plus a guard so both sides can't collapse to empty and match vacuously.

Green: sequential (1630) · --parallel --simple --strict (1589) · fork budget · make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅. CHANGELOG updated under Fixed.

[Chemaclass]

🤔 Background

Closes#959

Under --parallel, each test file's results were bucketed into a directory named by the file's basename:

local test_suite_base="${test_file##*/}"local test_suite_dir="${TEMP_DIR_PARALLEL_TEST_SUITE}/${test_suite_base%.sh}"

Two files sharing a basename in different directories shared one bucket. The per-suite ordinal is unique per suite, not globally, so <tmp>/foo_test/1.result was written twice and the first write was lost. The run stayed green and simply reported fewer tests.

Same defect as #923, which fixed exactly this in build.sh: a basename isn't an identity when directories can repeat one.

📉 How it was found, and why it matters to users

Mirroring tests/unit/ onto the src/ module layout (#957/#960) produced three duplicate basenames and cost 52 tests under --parallel:

	--parallel --simple --strict
with 3 duplicate basenames	1536
after renaming them unique	1588

A tests/ tree mirroring a src/ tree is what the docs encourage, so any user project doing it is exposed — and a failing test can vanish while CI stays green.

💡 Changes

• The derivation folds the whole path instead of taking the basename.
• Both sites that computed it — runner/result.sh and the pre-create in runner/exec.sh (perf(runner): stop forking per test when publishing parallel results #813) — now call one shared helper, so they cannot drift apart.
• It stays pure parameter expansion: this runs once per test in every parallel worker and the path must remain fork-free. The fork-budget acceptance tests pass unchanged (11/11).

Nothing parses the directory name — aggregate_parallel_results globs whatever subdirectories exist — so the rename is invisible downstream.

✅ Verification

TDD: the new acceptance test in bashunit_parallel_consistency_test.sh runs two same-named fixture files from different directories in both modes and compares counts. Verified red without the fix (sequential 2, parallel 1) and green with it.

It asserts on extracted counts rather than the rendered summary — the parallel spinner leaves control bytes on that line — plus a guard so both sides can't collapse to empty and match vacuously.

Green: sequential (1630) · --parallel --simple --strict (1589) · fork budget · make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅. CHANGELOG updated under Fixed.

[Chemaclass]

🤔 Background

Closes#959

Under --parallel, each test file's results were bucketed into a directory named by the file's basename:

local test_suite_base="${test_file##*/}"local test_suite_dir="${TEMP_DIR_PARALLEL_TEST_SUITE}/${test_suite_base%.sh}"

Two files sharing a basename in different directories shared one bucket. The per-suite ordinal is unique per suite, not globally, so <tmp>/foo_test/1.result was written twice and the first write was lost. The run stayed green and simply reported fewer tests.

Same defect as #923, which fixed exactly this in build.sh: a basename isn't an identity when directories can repeat one.

📉 How it was found, and why it matters to users

Mirroring tests/unit/ onto the src/ module layout (#957/#960) produced three duplicate basenames and cost 52 tests under --parallel:

	--parallel --simple --strict
with 3 duplicate basenames	1536
after renaming them unique	1588

A tests/ tree mirroring a src/ tree is what the docs encourage, so any user project doing it is exposed — and a failing test can vanish while CI stays green.

💡 Changes

• The derivation folds the whole path instead of taking the basename.
• Both sites that computed it — runner/result.sh and the pre-create in runner/exec.sh (perf(runner): stop forking per test when publishing parallel results #813) — now call one shared helper, so they cannot drift apart.
• It stays pure parameter expansion: this runs once per test in every parallel worker and the path must remain fork-free. The fork-budget acceptance tests pass unchanged (11/11).

Nothing parses the directory name — aggregate_parallel_results globs whatever subdirectories exist — so the rename is invisible downstream.

✅ Verification

TDD: the new acceptance test in bashunit_parallel_consistency_test.sh runs two same-named fixture files from different directories in both modes and compares counts. Verified red without the fix (sequential 2, parallel 1) and green with it.

It asserts on extracted counts rather than the rendered summary — the parallel spinner leaves control bytes on that line — plus a guard so both sides can't collapse to empty and match vacuously.

Green: sequential (1630) · --parallel --simple --strict (1589) · fork budget · make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅. CHANGELOG updated under Fixed.

[Chemaclass]

🤔 Background

Closes#959

Under --parallel, each test file's results were bucketed into a directory named by the file's basename:

local test_suite_base="${test_file##*/}"local test_suite_dir="${TEMP_DIR_PARALLEL_TEST_SUITE}/${test_suite_base%.sh}"

Two files sharing a basename in different directories shared one bucket. The per-suite ordinal is unique per suite, not globally, so <tmp>/foo_test/1.result was written twice and the first write was lost. The run stayed green and simply reported fewer tests.

Same defect as #923, which fixed exactly this in build.sh: a basename isn't an identity when directories can repeat one.

📉 How it was found, and why it matters to users

Mirroring tests/unit/ onto the src/ module layout (#957/#960) produced three duplicate basenames and cost 52 tests under --parallel:

	--parallel --simple --strict
with 3 duplicate basenames	1536
after renaming them unique	1588

A tests/ tree mirroring a src/ tree is what the docs encourage, so any user project doing it is exposed — and a failing test can vanish while CI stays green.

💡 Changes

• The derivation folds the whole path instead of taking the basename.
• Both sites that computed it — runner/result.sh and the pre-create in runner/exec.sh (perf(runner): stop forking per test when publishing parallel results #813) — now call one shared helper, so they cannot drift apart.
• It stays pure parameter expansion: this runs once per test in every parallel worker and the path must remain fork-free. The fork-budget acceptance tests pass unchanged (11/11).

Nothing parses the directory name — aggregate_parallel_results globs whatever subdirectories exist — so the rename is invisible downstream.

✅ Verification

TDD: the new acceptance test in bashunit_parallel_consistency_test.sh runs two same-named fixture files from different directories in both modes and compares counts. Verified red without the fix (sequential 2, parallel 1) and green with it.

It asserts on extracted counts rather than the rendered summary — the parallel spinner leaves control bytes on that line — plus a guard so both sides can't collapse to empty and match vacuously.

Green: sequential (1630) · --parallel --simple --strict (1589) · fork budget · make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅. CHANGELOG updated under Fixed.

[Chemaclass]

🤔 Background

Closes#959

Under --parallel, each test file's results were bucketed into a directory named by the file's basename:

local test_suite_base="${test_file##*/}"local test_suite_dir="${TEMP_DIR_PARALLEL_TEST_SUITE}/${test_suite_base%.sh}"

Two files sharing a basename in different directories shared one bucket. The per-suite ordinal is unique per suite, not globally, so <tmp>/foo_test/1.result was written twice and the first write was lost. The run stayed green and simply reported fewer tests.

Same defect as #923, which fixed exactly this in build.sh: a basename isn't an identity when directories can repeat one.

📉 How it was found, and why it matters to users

Mirroring tests/unit/ onto the src/ module layout (#957/#960) produced three duplicate basenames and cost 52 tests under --parallel:

	--parallel --simple --strict
with 3 duplicate basenames	1536
after renaming them unique	1588

A tests/ tree mirroring a src/ tree is what the docs encourage, so any user project doing it is exposed — and a failing test can vanish while CI stays green.

💡 Changes

• The derivation folds the whole path instead of taking the basename.
• Both sites that computed it — runner/result.sh and the pre-create in runner/exec.sh (perf(runner): stop forking per test when publishing parallel results #813) — now call one shared helper, so they cannot drift apart.
• It stays pure parameter expansion: this runs once per test in every parallel worker and the path must remain fork-free. The fork-budget acceptance tests pass unchanged (11/11).

Nothing parses the directory name — aggregate_parallel_results globs whatever subdirectories exist — so the rename is invisible downstream.

✅ Verification

TDD: the new acceptance test in bashunit_parallel_consistency_test.sh runs two same-named fixture files from different directories in both modes and compares counts. Verified red without the fix (sequential 2, parallel 1) and green with it.

It asserts on extracted counts rather than the rendered summary — the parallel spinner leaves control bytes on that line — plus a guard so both sides can't collapse to empty and match vacuously.

Green: sequential (1630) · --parallel --simple --strict (1589) · fork budget · make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅. CHANGELOG updated under Fixed.

[Chemaclass]

🤔 Background

Closes#959

Under --parallel, each test file's results were bucketed into a directory named by the file's basename:

local test_suite_base="${test_file##*/}"local test_suite_dir="${TEMP_DIR_PARALLEL_TEST_SUITE}/${test_suite_base%.sh}"

Two files sharing a basename in different directories shared one bucket. The per-suite ordinal is unique per suite, not globally, so <tmp>/foo_test/1.result was written twice and the first write was lost. The run stayed green and simply reported fewer tests.

Same defect as #923, which fixed exactly this in build.sh: a basename isn't an identity when directories can repeat one.

📉 How it was found, and why it matters to users

Mirroring tests/unit/ onto the src/ module layout (#957/#960) produced three duplicate basenames and cost 52 tests under --parallel:

	--parallel --simple --strict
with 3 duplicate basenames	1536
after renaming them unique	1588

A tests/ tree mirroring a src/ tree is what the docs encourage, so any user project doing it is exposed — and a failing test can vanish while CI stays green.

💡 Changes

• The derivation folds the whole path instead of taking the basename.
• Both sites that computed it — runner/result.sh and the pre-create in runner/exec.sh (perf(runner): stop forking per test when publishing parallel results #813) — now call one shared helper, so they cannot drift apart.
• It stays pure parameter expansion: this runs once per test in every parallel worker and the path must remain fork-free. The fork-budget acceptance tests pass unchanged (11/11).

Nothing parses the directory name — aggregate_parallel_results globs whatever subdirectories exist — so the rename is invisible downstream.

✅ Verification

TDD: the new acceptance test in bashunit_parallel_consistency_test.sh runs two same-named fixture files from different directories in both modes and compares counts. Verified red without the fix (sequential 2, parallel 1) and green with it.

It asserts on extracted counts rather than the rendered summary — the parallel spinner leaves control bytes on that line — plus a guard so both sides can't collapse to empty and match vacuously.

Green: sequential (1630) · --parallel --simple --strict (1589) · fork budget · make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅. CHANGELOG updated under Fixed.

[Chemaclass]

🤔 Background

Closes#959

Under --parallel, each test file's results were bucketed into a directory named by the file's basename:

local test_suite_base="${test_file##*/}"local test_suite_dir="${TEMP_DIR_PARALLEL_TEST_SUITE}/${test_suite_base%.sh}"

Two files sharing a basename in different directories shared one bucket. The per-suite ordinal is unique per suite, not globally, so <tmp>/foo_test/1.result was written twice and the first write was lost. The run stayed green and simply reported fewer tests.

Same defect as #923, which fixed exactly this in build.sh: a basename isn't an identity when directories can repeat one.

📉 How it was found, and why it matters to users

Mirroring tests/unit/ onto the src/ module layout (#957/#960) produced three duplicate basenames and cost 52 tests under --parallel:

	--parallel --simple --strict
with 3 duplicate basenames	1536
after renaming them unique	1588

A tests/ tree mirroring a src/ tree is what the docs encourage, so any user project doing it is exposed — and a failing test can vanish while CI stays green.

💡 Changes

• The derivation folds the whole path instead of taking the basename.
• Both sites that computed it — runner/result.sh and the pre-create in runner/exec.sh (perf(runner): stop forking per test when publishing parallel results #813) — now call one shared helper, so they cannot drift apart.
• It stays pure parameter expansion: this runs once per test in every parallel worker and the path must remain fork-free. The fork-budget acceptance tests pass unchanged (11/11).

Nothing parses the directory name — aggregate_parallel_results globs whatever subdirectories exist — so the rename is invisible downstream.

✅ Verification

TDD: the new acceptance test in bashunit_parallel_consistency_test.sh runs two same-named fixture files from different directories in both modes and compares counts. Verified red without the fix (sequential 2, parallel 1) and green with it.

It asserts on extracted counts rather than the rendered summary — the parallel spinner leaves control bytes on that line — plus a guard so both sides can't collapse to empty and match vacuously.

Green: sequential (1630) · --parallel --simple --strict (1589) · fork budget · make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅. CHANGELOG updated under Fixed.