fix(ci): unit and acceptance shards matched nothing and silently ran everything

[Chemaclass]

🤔 Background

The macOS and Windows matrices sliced tests/unit/ with alphabetical globs — tests/unit/[a-b]*_test.sh and four siblings. #960 moved every unit test into a per-module subdirectory, so all five stopped matching any file.

That didn't fail. bashunit answers an argument matching no file by falling back to its default path — the whole suite. So all ten macOS and Windows unit legs have been running all 1652 tests instead of their fifth, on every run since #960, green throughout. Nothing was broken, so nothing pointed at it.

The acceptance buckets are the same design and had rotted the other way, costing coverage instead of time: tests/acceptance/worker_stderr_test.sh matched none of the three patterns, so it has never run in CI since #891 added it.

💡 Changes

Both are now --shard i/n <dir>, using the flag bashunit already ships. It cannot select nothing, needs no maintenance when files move, and partitions exactly:

Suite	Shards	Sum	Whole
tests/unit	229 + 166 + 219 + 277 + 326	1217	1217 ✅
tests/acceptance	128 + 128 + 98	354	354 ✅

No gaps, no overlap.

🧪 The guard

A new contract test pins the property that mattered in both cases: a path CI runs must resolve to something. Mutation-tested three ways — restoring a dead glob, pointing a shard at a missing directory, renaming the workflow away — each turns it red.

📋 Note

Sharding by file leaves the legs uneven (166–326 tests). Balancing them is a separate question from them running the right tests at all.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1652 sequential / 1611 parallel-simple-strict.

[Chemaclass]

🤔 Background

The macOS and Windows matrices sliced tests/unit/ with alphabetical globs — tests/unit/[a-b]*_test.sh and four siblings. #960 moved every unit test into a per-module subdirectory, so all five stopped matching any file.

That didn't fail. bashunit answers an argument matching no file by falling back to its default path — the whole suite. So all ten macOS and Windows unit legs have been running all 1652 tests instead of their fifth, on every run since #960, green throughout. Nothing was broken, so nothing pointed at it.

The acceptance buckets are the same design and had rotted the other way, costing coverage instead of time: tests/acceptance/worker_stderr_test.sh matched none of the three patterns, so it has never run in CI since #891 added it.

💡 Changes

Both are now --shard i/n <dir>, using the flag bashunit already ships. It cannot select nothing, needs no maintenance when files move, and partitions exactly:

Suite	Shards	Sum	Whole
tests/unit	229 + 166 + 219 + 277 + 326	1217	1217 ✅
tests/acceptance	128 + 128 + 98	354	354 ✅

No gaps, no overlap.

🧪 The guard

A new contract test pins the property that mattered in both cases: a path CI runs must resolve to something. Mutation-tested three ways — restoring a dead glob, pointing a shard at a missing directory, renaming the workflow away — each turns it red.

📋 Note

Sharding by file leaves the legs uneven (166–326 tests). Balancing them is a separate question from them running the right tests at all.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1652 sequential / 1611 parallel-simple-strict.

[Chemaclass]

🤔 Background

The macOS and Windows matrices sliced tests/unit/ with alphabetical globs — tests/unit/[a-b]*_test.sh and four siblings. #960 moved every unit test into a per-module subdirectory, so all five stopped matching any file.

That didn't fail. bashunit answers an argument matching no file by falling back to its default path — the whole suite. So all ten macOS and Windows unit legs have been running all 1652 tests instead of their fifth, on every run since #960, green throughout. Nothing was broken, so nothing pointed at it.

The acceptance buckets are the same design and had rotted the other way, costing coverage instead of time: tests/acceptance/worker_stderr_test.sh matched none of the three patterns, so it has never run in CI since #891 added it.

💡 Changes

Both are now --shard i/n <dir>, using the flag bashunit already ships. It cannot select nothing, needs no maintenance when files move, and partitions exactly:

Suite	Shards	Sum	Whole
tests/unit	229 + 166 + 219 + 277 + 326	1217	1217 ✅
tests/acceptance	128 + 128 + 98	354	354 ✅

No gaps, no overlap.

🧪 The guard

A new contract test pins the property that mattered in both cases: a path CI runs must resolve to something. Mutation-tested three ways — restoring a dead glob, pointing a shard at a missing directory, renaming the workflow away — each turns it red.

📋 Note

Sharding by file leaves the legs uneven (166–326 tests). Balancing them is a separate question from them running the right tests at all.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1652 sequential / 1611 parallel-simple-strict.

[Chemaclass]

🤔 Background

The macOS and Windows matrices sliced tests/unit/ with alphabetical globs — tests/unit/[a-b]*_test.sh and four siblings. #960 moved every unit test into a per-module subdirectory, so all five stopped matching any file.

That didn't fail. bashunit answers an argument matching no file by falling back to its default path — the whole suite. So all ten macOS and Windows unit legs have been running all 1652 tests instead of their fifth, on every run since #960, green throughout. Nothing was broken, so nothing pointed at it.

The acceptance buckets are the same design and had rotted the other way, costing coverage instead of time: tests/acceptance/worker_stderr_test.sh matched none of the three patterns, so it has never run in CI since #891 added it.

💡 Changes

Both are now --shard i/n <dir>, using the flag bashunit already ships. It cannot select nothing, needs no maintenance when files move, and partitions exactly:

Suite	Shards	Sum	Whole
tests/unit	229 + 166 + 219 + 277 + 326	1217	1217 ✅
tests/acceptance	128 + 128 + 98	354	354 ✅

No gaps, no overlap.

🧪 The guard

A new contract test pins the property that mattered in both cases: a path CI runs must resolve to something. Mutation-tested three ways — restoring a dead glob, pointing a shard at a missing directory, renaming the workflow away — each turns it red.

📋 Note

Sharding by file leaves the legs uneven (166–326 tests). Balancing them is a separate question from them running the right tests at all.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1652 sequential / 1611 parallel-simple-strict.

[Chemaclass]

🤔 Background

The macOS and Windows matrices sliced tests/unit/ with alphabetical globs — tests/unit/[a-b]*_test.sh and four siblings. #960 moved every unit test into a per-module subdirectory, so all five stopped matching any file.

That didn't fail. bashunit answers an argument matching no file by falling back to its default path — the whole suite. So all ten macOS and Windows unit legs have been running all 1652 tests instead of their fifth, on every run since #960, green throughout. Nothing was broken, so nothing pointed at it.

The acceptance buckets are the same design and had rotted the other way, costing coverage instead of time: tests/acceptance/worker_stderr_test.sh matched none of the three patterns, so it has never run in CI since #891 added it.

💡 Changes

Both are now --shard i/n <dir>, using the flag bashunit already ships. It cannot select nothing, needs no maintenance when files move, and partitions exactly:

Suite	Shards	Sum	Whole
tests/unit	229 + 166 + 219 + 277 + 326	1217	1217 ✅
tests/acceptance	128 + 128 + 98	354	354 ✅

No gaps, no overlap.

🧪 The guard

A new contract test pins the property that mattered in both cases: a path CI runs must resolve to something. Mutation-tested three ways — restoring a dead glob, pointing a shard at a missing directory, renaming the workflow away — each turns it red.

📋 Note

Sharding by file leaves the legs uneven (166–326 tests). Balancing them is a separate question from them running the right tests at all.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1652 sequential / 1611 parallel-simple-strict.

[Chemaclass]

🤔 Background

The macOS and Windows matrices sliced tests/unit/ with alphabetical globs — tests/unit/[a-b]*_test.sh and four siblings. #960 moved every unit test into a per-module subdirectory, so all five stopped matching any file.

That didn't fail. bashunit answers an argument matching no file by falling back to its default path — the whole suite. So all ten macOS and Windows unit legs have been running all 1652 tests instead of their fifth, on every run since #960, green throughout. Nothing was broken, so nothing pointed at it.

The acceptance buckets are the same design and had rotted the other way, costing coverage instead of time: tests/acceptance/worker_stderr_test.sh matched none of the three patterns, so it has never run in CI since #891 added it.

💡 Changes

Both are now --shard i/n <dir>, using the flag bashunit already ships. It cannot select nothing, needs no maintenance when files move, and partitions exactly:

Suite	Shards	Sum	Whole
tests/unit	229 + 166 + 219 + 277 + 326	1217	1217 ✅
tests/acceptance	128 + 128 + 98	354	354 ✅

No gaps, no overlap.

🧪 The guard

A new contract test pins the property that mattered in both cases: a path CI runs must resolve to something. Mutation-tested three ways — restoring a dead glob, pointing a shard at a missing directory, renaming the workflow away — each turns it red.

📋 Note

Sharding by file leaves the legs uneven (166–326 tests). Balancing them is a separate question from them running the right tests at all.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1652 sequential / 1611 parallel-simple-strict.

[Chemaclass]

🤔 Background

The macOS and Windows matrices sliced tests/unit/ with alphabetical globs — tests/unit/[a-b]*_test.sh and four siblings. #960 moved every unit test into a per-module subdirectory, so all five stopped matching any file.

That didn't fail. bashunit answers an argument matching no file by falling back to its default path — the whole suite. So all ten macOS and Windows unit legs have been running all 1652 tests instead of their fifth, on every run since #960, green throughout. Nothing was broken, so nothing pointed at it.

The acceptance buckets are the same design and had rotted the other way, costing coverage instead of time: tests/acceptance/worker_stderr_test.sh matched none of the three patterns, so it has never run in CI since #891 added it.

💡 Changes

Both are now --shard i/n <dir>, using the flag bashunit already ships. It cannot select nothing, needs no maintenance when files move, and partitions exactly:

Suite	Shards	Sum	Whole
tests/unit	229 + 166 + 219 + 277 + 326	1217	1217 ✅
tests/acceptance	128 + 128 + 98	354	354 ✅

No gaps, no overlap.

🧪 The guard

A new contract test pins the property that mattered in both cases: a path CI runs must resolve to something. Mutation-tested three ways — restoring a dead glob, pointing a shard at a missing directory, renaming the workflow away — each turns it red.

📋 Note

Sharding by file leaves the legs uneven (166–326 tests). Balancing them is a separate question from them running the right tests at all.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1652 sequential / 1611 parallel-simple-strict.

[Chemaclass]

🤔 Background

The macOS and Windows matrices sliced tests/unit/ with alphabetical globs — tests/unit/[a-b]*_test.sh and four siblings. #960 moved every unit test into a per-module subdirectory, so all five stopped matching any file.

That didn't fail. bashunit answers an argument matching no file by falling back to its default path — the whole suite. So all ten macOS and Windows unit legs have been running all 1652 tests instead of their fifth, on every run since #960, green throughout. Nothing was broken, so nothing pointed at it.

The acceptance buckets are the same design and had rotted the other way, costing coverage instead of time: tests/acceptance/worker_stderr_test.sh matched none of the three patterns, so it has never run in CI since #891 added it.

💡 Changes

Both are now --shard i/n <dir>, using the flag bashunit already ships. It cannot select nothing, needs no maintenance when files move, and partitions exactly:

Suite	Shards	Sum	Whole
tests/unit	229 + 166 + 219 + 277 + 326	1217	1217 ✅
tests/acceptance	128 + 128 + 98	354	354 ✅

No gaps, no overlap.

🧪 The guard

A new contract test pins the property that mattered in both cases: a path CI runs must resolve to something. Mutation-tested three ways — restoring a dead glob, pointing a shard at a missing directory, renaming the workflow away — each turns it red.

📋 Note

Sharding by file leaves the legs uneven (166–326 tests). Balancing them is a separate question from them running the right tests at all.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1652 sequential / 1611 parallel-simple-strict.