docs(rules): record where per-test time actually goes by Chemaclass · Pull Request #1190 · TypedDevs/bashunit

Chemaclass · 2026-08-13T23:51:58Z

🤔 Background

Related #1189

shape	wall	per unit
1 test, 500 `assert_same`	~73 ms	0.15 ms per assertion
500 tests, 1 assertion each	~3.9 s	7.8 ms per test

A test costs ~50x an assertion, which is why #801–#851 targeted per-test forks
and not the assertion layer.

💡 Changes

Records the profile in .claude/rules/perf-fork-budget.md, including three
explanations that were checked and ruled out:

Not forks — a 30-binary PATH-shim census counts 12 forks for 500 tests
Not the capture subshell — a bare $( ) is ~0.46 ms, ~6% of the 7.8 ms
Not quadratic — 7.17 ms/test at 100 tests vs 8.06 ms at 1000, so perf(runner): unset each file's test functions after processing #830 holds

No optimization was found in this pass. The numbers are the deliverable:
they rule out assertion micro-optimization, give a regression signal for #830,
and stop the next attempt re-deriving them.

Measured: a test costs ~7.8ms and an assertion ~0.15ms, so a test is worth about 50 assertions. That is why #801-#851 went after per-test forks and not the assertion layer, and it says assertion micro-optimization is not worth doing. Three explanations checked and ruled out, so nobody re-checks them: a 30-binary PATH-shim census counts 12 forks for 500 tests, a bare $( ) is ~0.46ms (about 6% of the per-test cost), and per-test cost is 7.17ms at 100 tests against 8.06ms at 1000 -- +12% over a 10x range, so #830 still holds. No win found in this pass. The numbers are the deliverable: they give a regression signal for #830 and save the next attempt from re-deriving them. Closes#1189

Chemaclass · 2026-08-13T23:51:58Z

🤔 Background

Related #1189

shape	wall	per unit
1 test, 500 `assert_same`	~73 ms	0.15 ms per assertion
500 tests, 1 assertion each	~3.9 s	7.8 ms per test

A test costs ~50x an assertion, which is why #801–#851 targeted per-test forks
and not the assertion layer.

💡 Changes

Records the profile in .claude/rules/perf-fork-budget.md, including three
explanations that were checked and ruled out:

Not forks — a 30-binary PATH-shim census counts 12 forks for 500 tests
Not the capture subshell — a bare $( ) is ~0.46 ms, ~6% of the 7.8 ms
Not quadratic — 7.17 ms/test at 100 tests vs 8.06 ms at 1000, so perf(runner): unset each file's test functions after processing #830 holds

No optimization was found in this pass. The numbers are the deliverable:
they rule out assertion micro-optimization, give a regression signal for #830,
and stop the next attempt re-deriving them.

Measured: a test costs ~7.8ms and an assertion ~0.15ms, so a test is worth about 50 assertions. That is why #801-#851 went after per-test forks and not the assertion layer, and it says assertion micro-optimization is not worth doing. Three explanations checked and ruled out, so nobody re-checks them: a 30-binary PATH-shim census counts 12 forks for 500 tests, a bare $( ) is ~0.46ms (about 6% of the per-test cost), and per-test cost is 7.17ms at 100 tests against 8.06ms at 1000 -- +12% over a 10x range, so #830 still holds. No win found in this pass. The numbers are the deliverable: they give a regression signal for #830 and save the next attempt from re-deriving them. Closes#1189

Chemaclass · 2026-08-13T23:51:58Z

🤔 Background

Related #1189

shape	wall	per unit
1 test, 500 `assert_same`	~73 ms	0.15 ms per assertion
500 tests, 1 assertion each	~3.9 s	7.8 ms per test

A test costs ~50x an assertion, which is why #801–#851 targeted per-test forks
and not the assertion layer.

💡 Changes

Records the profile in .claude/rules/perf-fork-budget.md, including three
explanations that were checked and ruled out:

Not forks — a 30-binary PATH-shim census counts 12 forks for 500 tests
Not the capture subshell — a bare $( ) is ~0.46 ms, ~6% of the 7.8 ms
Not quadratic — 7.17 ms/test at 100 tests vs 8.06 ms at 1000, so perf(runner): unset each file's test functions after processing #830 holds

No optimization was found in this pass. The numbers are the deliverable:
they rule out assertion micro-optimization, give a regression signal for #830,
and stop the next attempt re-deriving them.

Measured: a test costs ~7.8ms and an assertion ~0.15ms, so a test is worth about 50 assertions. That is why #801-#851 went after per-test forks and not the assertion layer, and it says assertion micro-optimization is not worth doing. Three explanations checked and ruled out, so nobody re-checks them: a 30-binary PATH-shim census counts 12 forks for 500 tests, a bare $( ) is ~0.46ms (about 6% of the per-test cost), and per-test cost is 7.17ms at 100 tests against 8.06ms at 1000 -- +12% over a 10x range, so #830 still holds. No win found in this pass. The numbers are the deliverable: they give a regression signal for #830 and save the next attempt from re-deriving them. Closes#1189

Chemaclass · 2026-08-13T23:51:58Z

🤔 Background

Related #1189

shape	wall	per unit
1 test, 500 `assert_same`	~73 ms	0.15 ms per assertion
500 tests, 1 assertion each	~3.9 s	7.8 ms per test

A test costs ~50x an assertion, which is why #801–#851 targeted per-test forks
and not the assertion layer.

💡 Changes

Records the profile in .claude/rules/perf-fork-budget.md, including three
explanations that were checked and ruled out:

Not forks — a 30-binary PATH-shim census counts 12 forks for 500 tests
Not the capture subshell — a bare $( ) is ~0.46 ms, ~6% of the 7.8 ms
Not quadratic — 7.17 ms/test at 100 tests vs 8.06 ms at 1000, so perf(runner): unset each file's test functions after processing #830 holds

No optimization was found in this pass. The numbers are the deliverable:
they rule out assertion micro-optimization, give a regression signal for #830,
and stop the next attempt re-deriving them.

Measured: a test costs ~7.8ms and an assertion ~0.15ms, so a test is worth about 50 assertions. That is why #801-#851 went after per-test forks and not the assertion layer, and it says assertion micro-optimization is not worth doing. Three explanations checked and ruled out, so nobody re-checks them: a 30-binary PATH-shim census counts 12 forks for 500 tests, a bare $( ) is ~0.46ms (about 6% of the per-test cost), and per-test cost is 7.17ms at 100 tests against 8.06ms at 1000 -- +12% over a 10x range, so #830 still holds. No win found in this pass. The numbers are the deliverable: they give a regression signal for #830 and save the next attempt from re-deriving them. Closes#1189

Chemaclass · 2026-08-13T23:51:58Z

🤔 Background

Related #1189

shape	wall	per unit
1 test, 500 `assert_same`	~73 ms	0.15 ms per assertion
500 tests, 1 assertion each	~3.9 s	7.8 ms per test

A test costs ~50x an assertion, which is why #801–#851 targeted per-test forks
and not the assertion layer.

💡 Changes

Records the profile in .claude/rules/perf-fork-budget.md, including three
explanations that were checked and ruled out:

Not forks — a 30-binary PATH-shim census counts 12 forks for 500 tests
Not the capture subshell — a bare $( ) is ~0.46 ms, ~6% of the 7.8 ms
Not quadratic — 7.17 ms/test at 100 tests vs 8.06 ms at 1000, so perf(runner): unset each file's test functions after processing #830 holds

No optimization was found in this pass. The numbers are the deliverable:
they rule out assertion micro-optimization, give a regression signal for #830,
and stop the next attempt re-deriving them.

Measured: a test costs ~7.8ms and an assertion ~0.15ms, so a test is worth about 50 assertions. That is why #801-#851 went after per-test forks and not the assertion layer, and it says assertion micro-optimization is not worth doing. Three explanations checked and ruled out, so nobody re-checks them: a 30-binary PATH-shim census counts 12 forks for 500 tests, a bare $( ) is ~0.46ms (about 6% of the per-test cost), and per-test cost is 7.17ms at 100 tests against 8.06ms at 1000 -- +12% over a 10x range, so #830 still holds. No win found in this pass. The numbers are the deliverable: they give a regression signal for #830 and save the next attempt from re-deriving them. Closes#1189

Chemaclass · 2026-08-13T23:51:58Z

🤔 Background

Related #1189

shape	wall	per unit
1 test, 500 `assert_same`	~73 ms	0.15 ms per assertion
500 tests, 1 assertion each	~3.9 s	7.8 ms per test

A test costs ~50x an assertion, which is why #801–#851 targeted per-test forks
and not the assertion layer.

💡 Changes

Records the profile in .claude/rules/perf-fork-budget.md, including three
explanations that were checked and ruled out:

Not forks — a 30-binary PATH-shim census counts 12 forks for 500 tests
Not the capture subshell — a bare $( ) is ~0.46 ms, ~6% of the 7.8 ms
Not quadratic — 7.17 ms/test at 100 tests vs 8.06 ms at 1000, so perf(runner): unset each file's test functions after processing #830 holds

No optimization was found in this pass. The numbers are the deliverable:
they rule out assertion micro-optimization, give a regression signal for #830,
and stop the next attempt re-deriving them.

Measured: a test costs ~7.8ms and an assertion ~0.15ms, so a test is worth about 50 assertions. That is why #801-#851 went after per-test forks and not the assertion layer, and it says assertion micro-optimization is not worth doing. Three explanations checked and ruled out, so nobody re-checks them: a 30-binary PATH-shim census counts 12 forks for 500 tests, a bare $( ) is ~0.46ms (about 6% of the per-test cost), and per-test cost is 7.17ms at 100 tests against 8.06ms at 1000 -- +12% over a 10x range, so #830 still holds. No win found in this pass. The numbers are the deliverable: they give a regression signal for #830 and save the next attempt from re-deriving them. Closes#1189

Chemaclass · 2026-08-13T23:51:58Z

🤔 Background

Related #1189

shape	wall	per unit
1 test, 500 `assert_same`	~73 ms	0.15 ms per assertion
500 tests, 1 assertion each	~3.9 s	7.8 ms per test

A test costs ~50x an assertion, which is why #801–#851 targeted per-test forks
and not the assertion layer.

💡 Changes

Records the profile in .claude/rules/perf-fork-budget.md, including three
explanations that were checked and ruled out:

Not forks — a 30-binary PATH-shim census counts 12 forks for 500 tests
Not the capture subshell — a bare $( ) is ~0.46 ms, ~6% of the 7.8 ms
Not quadratic — 7.17 ms/test at 100 tests vs 8.06 ms at 1000, so perf(runner): unset each file's test functions after processing #830 holds

No optimization was found in this pass. The numbers are the deliverable:
they rule out assertion micro-optimization, give a regression signal for #830,
and stop the next attempt re-deriving them.

Measured: a test costs ~7.8ms and an assertion ~0.15ms, so a test is worth about 50 assertions. That is why #801-#851 went after per-test forks and not the assertion layer, and it says assertion micro-optimization is not worth doing. Three explanations checked and ruled out, so nobody re-checks them: a 30-binary PATH-shim census counts 12 forks for 500 tests, a bare $( ) is ~0.46ms (about 6% of the per-test cost), and per-test cost is 7.17ms at 100 tests against 8.06ms at 1000 -- +12% over a 10x range, so #830 still holds. No win found in this pass. The numbers are the deliverable: they give a regression signal for #830 and save the next attempt from re-deriving them. Closes#1189

Chemaclass · 2026-08-13T23:51:58Z

🤔 Background

Related #1189

shape	wall	per unit
1 test, 500 `assert_same`	~73 ms	0.15 ms per assertion
500 tests, 1 assertion each	~3.9 s	7.8 ms per test

A test costs ~50x an assertion, which is why #801–#851 targeted per-test forks
and not the assertion layer.

💡 Changes

Records the profile in .claude/rules/perf-fork-budget.md, including three
explanations that were checked and ruled out:

Not forks — a 30-binary PATH-shim census counts 12 forks for 500 tests
Not the capture subshell — a bare $( ) is ~0.46 ms, ~6% of the 7.8 ms
Not quadratic — 7.17 ms/test at 100 tests vs 8.06 ms at 1000, so perf(runner): unset each file's test functions after processing #830 holds

No optimization was found in this pass. The numbers are the deliverable:
they rule out assertion micro-optimization, give a regression signal for #830,
and stop the next attempt re-deriving them.

Measured: a test costs ~7.8ms and an assertion ~0.15ms, so a test is worth about 50 assertions. That is why #801-#851 went after per-test forks and not the assertion layer, and it says assertion micro-optimization is not worth doing. Three explanations checked and ruled out, so nobody re-checks them: a 30-binary PATH-shim census counts 12 forks for 500 tests, a bare $( ) is ~0.46ms (about 6% of the per-test cost), and per-test cost is 7.17ms at 100 tests against 8.06ms at 1000 -- +12% over a 10x range, so #830 still holds. No win found in this pass. The numbers are the deliverable: they give a regression signal for #830 and save the next attempt from re-deriving them. Closes#1189

Chemaclass added the documentation Improvements or additions to documentation label Aug 13, 2026

Chemaclass self-assigned this Aug 13, 2026

Chemaclass merged commit 63d6f7d into mainAug 13, 2026
6 checks passed

Chemaclass deleted the docs/1189-per-test-cost-profile branch August 13, 2026 23:56