test(state): close the assert-shaped hole in the layering contract

[Chemaclass]

🤔 Background

state_test.sh has banned state/ from calling console_results, console_header, parallel and runner since #868/#862. It never banned assert — and assert is now the direction the cycle runs in.

assert -> state is the expected, dominant edge: assertions record their own results. The reverse arrived with assert_once (#921): its marker shares state's per-test reset lifecycle, so state/context.sh calls assert::once_reset and state/counters.sh calls assert::once_is_absorbing. console/test_line.sh reaches into the same feature four more times.

That's a genuine two-node call-graph cycle — invisible to a walk of the source graph, because bash resolves calls at runtime. A full audit of that graph found nothing here.

💡 Changes

• The contract now allowlists exactly those two functions. The exception is deliberate and shipped; anything beyond it is the shape the file already bans. Mutation-tested: an added bashunit::assert::fail_with in state/ turns it red.
• A second test guards the greps themselves.refactor(state): split src/state.sh into a src/state/ module #946 lost a contract this way once: an unmatched glob is handed to grep as a literal, fails No such file, is swallowed by the || true these greps need, and reports a clean pass having scanned nothing. Globbing the module fixed the refactor(state): split src/state.sh into a src/state/ module #946 case but not a wholesale rename. Proven against a module path that doesn't exist.

📝 Three doc corrections, one to my own text

ADR-011 listed api/globals.sh running set -euo pipefail as a load-bearing ordering constraint. It isn't. The entrypoint sets the same options on its line 2, and system/dependencies.sh repeats them at module #2 — both before api/ (#4) loads. The line earns its place for anything sourcing that file alone, but moving api/ would not turn strict mode off. Struck through rather than deleted, correction dated.

ADR-011 also called "main/ sourced last" load-bearing. True as a description — sourcing main/ first was tried and the suite still passed, because every call resolves at runtime. Reworded.

architecture-map.md and src/api/index.sh both attributed assert_once to api/bashunit.sh. It lives in src/assert/once.sh.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1654 sequential / 1613 parallel-simple-strict — baseline + the 2 new contracts.

[Chemaclass]

🤔 Background

state_test.sh has banned state/ from calling console_results, console_header, parallel and runner since #868/#862. It never banned assert — and assert is now the direction the cycle runs in.

assert -> state is the expected, dominant edge: assertions record their own results. The reverse arrived with assert_once (#921): its marker shares state's per-test reset lifecycle, so state/context.sh calls assert::once_reset and state/counters.sh calls assert::once_is_absorbing. console/test_line.sh reaches into the same feature four more times.

That's a genuine two-node call-graph cycle — invisible to a walk of the source graph, because bash resolves calls at runtime. A full audit of that graph found nothing here.

💡 Changes

• The contract now allowlists exactly those two functions. The exception is deliberate and shipped; anything beyond it is the shape the file already bans. Mutation-tested: an added bashunit::assert::fail_with in state/ turns it red.
• A second test guards the greps themselves.refactor(state): split src/state.sh into a src/state/ module #946 lost a contract this way once: an unmatched glob is handed to grep as a literal, fails No such file, is swallowed by the || true these greps need, and reports a clean pass having scanned nothing. Globbing the module fixed the refactor(state): split src/state.sh into a src/state/ module #946 case but not a wholesale rename. Proven against a module path that doesn't exist.

📝 Three doc corrections, one to my own text

ADR-011 listed api/globals.sh running set -euo pipefail as a load-bearing ordering constraint. It isn't. The entrypoint sets the same options on its line 2, and system/dependencies.sh repeats them at module #2 — both before api/ (#4) loads. The line earns its place for anything sourcing that file alone, but moving api/ would not turn strict mode off. Struck through rather than deleted, correction dated.

ADR-011 also called "main/ sourced last" load-bearing. True as a description — sourcing main/ first was tried and the suite still passed, because every call resolves at runtime. Reworded.

architecture-map.md and src/api/index.sh both attributed assert_once to api/bashunit.sh. It lives in src/assert/once.sh.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1654 sequential / 1613 parallel-simple-strict — baseline + the 2 new contracts.

[Chemaclass]

🤔 Background

state_test.sh has banned state/ from calling console_results, console_header, parallel and runner since #868/#862. It never banned assert — and assert is now the direction the cycle runs in.

assert -> state is the expected, dominant edge: assertions record their own results. The reverse arrived with assert_once (#921): its marker shares state's per-test reset lifecycle, so state/context.sh calls assert::once_reset and state/counters.sh calls assert::once_is_absorbing. console/test_line.sh reaches into the same feature four more times.

That's a genuine two-node call-graph cycle — invisible to a walk of the source graph, because bash resolves calls at runtime. A full audit of that graph found nothing here.

💡 Changes

• The contract now allowlists exactly those two functions. The exception is deliberate and shipped; anything beyond it is the shape the file already bans. Mutation-tested: an added bashunit::assert::fail_with in state/ turns it red.
• A second test guards the greps themselves.refactor(state): split src/state.sh into a src/state/ module #946 lost a contract this way once: an unmatched glob is handed to grep as a literal, fails No such file, is swallowed by the || true these greps need, and reports a clean pass having scanned nothing. Globbing the module fixed the refactor(state): split src/state.sh into a src/state/ module #946 case but not a wholesale rename. Proven against a module path that doesn't exist.

📝 Three doc corrections, one to my own text

ADR-011 listed api/globals.sh running set -euo pipefail as a load-bearing ordering constraint. It isn't. The entrypoint sets the same options on its line 2, and system/dependencies.sh repeats them at module #2 — both before api/ (#4) loads. The line earns its place for anything sourcing that file alone, but moving api/ would not turn strict mode off. Struck through rather than deleted, correction dated.

ADR-011 also called "main/ sourced last" load-bearing. True as a description — sourcing main/ first was tried and the suite still passed, because every call resolves at runtime. Reworded.

architecture-map.md and src/api/index.sh both attributed assert_once to api/bashunit.sh. It lives in src/assert/once.sh.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1654 sequential / 1613 parallel-simple-strict — baseline + the 2 new contracts.

[Chemaclass]

🤔 Background

state_test.sh has banned state/ from calling console_results, console_header, parallel and runner since #868/#862. It never banned assert — and assert is now the direction the cycle runs in.

assert -> state is the expected, dominant edge: assertions record their own results. The reverse arrived with assert_once (#921): its marker shares state's per-test reset lifecycle, so state/context.sh calls assert::once_reset and state/counters.sh calls assert::once_is_absorbing. console/test_line.sh reaches into the same feature four more times.

That's a genuine two-node call-graph cycle — invisible to a walk of the source graph, because bash resolves calls at runtime. A full audit of that graph found nothing here.

💡 Changes

• The contract now allowlists exactly those two functions. The exception is deliberate and shipped; anything beyond it is the shape the file already bans. Mutation-tested: an added bashunit::assert::fail_with in state/ turns it red.
• A second test guards the greps themselves.refactor(state): split src/state.sh into a src/state/ module #946 lost a contract this way once: an unmatched glob is handed to grep as a literal, fails No such file, is swallowed by the || true these greps need, and reports a clean pass having scanned nothing. Globbing the module fixed the refactor(state): split src/state.sh into a src/state/ module #946 case but not a wholesale rename. Proven against a module path that doesn't exist.

📝 Three doc corrections, one to my own text

ADR-011 listed api/globals.sh running set -euo pipefail as a load-bearing ordering constraint. It isn't. The entrypoint sets the same options on its line 2, and system/dependencies.sh repeats them at module #2 — both before api/ (#4) loads. The line earns its place for anything sourcing that file alone, but moving api/ would not turn strict mode off. Struck through rather than deleted, correction dated.

ADR-011 also called "main/ sourced last" load-bearing. True as a description — sourcing main/ first was tried and the suite still passed, because every call resolves at runtime. Reworded.

architecture-map.md and src/api/index.sh both attributed assert_once to api/bashunit.sh. It lives in src/assert/once.sh.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1654 sequential / 1613 parallel-simple-strict — baseline + the 2 new contracts.

[Chemaclass]

🤔 Background

state_test.sh has banned state/ from calling console_results, console_header, parallel and runner since #868/#862. It never banned assert — and assert is now the direction the cycle runs in.

assert -> state is the expected, dominant edge: assertions record their own results. The reverse arrived with assert_once (#921): its marker shares state's per-test reset lifecycle, so state/context.sh calls assert::once_reset and state/counters.sh calls assert::once_is_absorbing. console/test_line.sh reaches into the same feature four more times.

That's a genuine two-node call-graph cycle — invisible to a walk of the source graph, because bash resolves calls at runtime. A full audit of that graph found nothing here.

💡 Changes

• The contract now allowlists exactly those two functions. The exception is deliberate and shipped; anything beyond it is the shape the file already bans. Mutation-tested: an added bashunit::assert::fail_with in state/ turns it red.
• A second test guards the greps themselves.refactor(state): split src/state.sh into a src/state/ module #946 lost a contract this way once: an unmatched glob is handed to grep as a literal, fails No such file, is swallowed by the || true these greps need, and reports a clean pass having scanned nothing. Globbing the module fixed the refactor(state): split src/state.sh into a src/state/ module #946 case but not a wholesale rename. Proven against a module path that doesn't exist.

📝 Three doc corrections, one to my own text

ADR-011 listed api/globals.sh running set -euo pipefail as a load-bearing ordering constraint. It isn't. The entrypoint sets the same options on its line 2, and system/dependencies.sh repeats them at module #2 — both before api/ (#4) loads. The line earns its place for anything sourcing that file alone, but moving api/ would not turn strict mode off. Struck through rather than deleted, correction dated.

ADR-011 also called "main/ sourced last" load-bearing. True as a description — sourcing main/ first was tried and the suite still passed, because every call resolves at runtime. Reworded.

architecture-map.md and src/api/index.sh both attributed assert_once to api/bashunit.sh. It lives in src/assert/once.sh.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1654 sequential / 1613 parallel-simple-strict — baseline + the 2 new contracts.

[Chemaclass]

🤔 Background

state_test.sh has banned state/ from calling console_results, console_header, parallel and runner since #868/#862. It never banned assert — and assert is now the direction the cycle runs in.

assert -> state is the expected, dominant edge: assertions record their own results. The reverse arrived with assert_once (#921): its marker shares state's per-test reset lifecycle, so state/context.sh calls assert::once_reset and state/counters.sh calls assert::once_is_absorbing. console/test_line.sh reaches into the same feature four more times.

That's a genuine two-node call-graph cycle — invisible to a walk of the source graph, because bash resolves calls at runtime. A full audit of that graph found nothing here.

💡 Changes

• The contract now allowlists exactly those two functions. The exception is deliberate and shipped; anything beyond it is the shape the file already bans. Mutation-tested: an added bashunit::assert::fail_with in state/ turns it red.
• A second test guards the greps themselves.refactor(state): split src/state.sh into a src/state/ module #946 lost a contract this way once: an unmatched glob is handed to grep as a literal, fails No such file, is swallowed by the || true these greps need, and reports a clean pass having scanned nothing. Globbing the module fixed the refactor(state): split src/state.sh into a src/state/ module #946 case but not a wholesale rename. Proven against a module path that doesn't exist.

📝 Three doc corrections, one to my own text

ADR-011 listed api/globals.sh running set -euo pipefail as a load-bearing ordering constraint. It isn't. The entrypoint sets the same options on its line 2, and system/dependencies.sh repeats them at module #2 — both before api/ (#4) loads. The line earns its place for anything sourcing that file alone, but moving api/ would not turn strict mode off. Struck through rather than deleted, correction dated.

ADR-011 also called "main/ sourced last" load-bearing. True as a description — sourcing main/ first was tried and the suite still passed, because every call resolves at runtime. Reworded.

architecture-map.md and src/api/index.sh both attributed assert_once to api/bashunit.sh. It lives in src/assert/once.sh.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1654 sequential / 1613 parallel-simple-strict — baseline + the 2 new contracts.

[Chemaclass]

🤔 Background

state_test.sh has banned state/ from calling console_results, console_header, parallel and runner since #868/#862. It never banned assert — and assert is now the direction the cycle runs in.

assert -> state is the expected, dominant edge: assertions record their own results. The reverse arrived with assert_once (#921): its marker shares state's per-test reset lifecycle, so state/context.sh calls assert::once_reset and state/counters.sh calls assert::once_is_absorbing. console/test_line.sh reaches into the same feature four more times.

That's a genuine two-node call-graph cycle — invisible to a walk of the source graph, because bash resolves calls at runtime. A full audit of that graph found nothing here.

💡 Changes

• The contract now allowlists exactly those two functions. The exception is deliberate and shipped; anything beyond it is the shape the file already bans. Mutation-tested: an added bashunit::assert::fail_with in state/ turns it red.
• A second test guards the greps themselves.refactor(state): split src/state.sh into a src/state/ module #946 lost a contract this way once: an unmatched glob is handed to grep as a literal, fails No such file, is swallowed by the || true these greps need, and reports a clean pass having scanned nothing. Globbing the module fixed the refactor(state): split src/state.sh into a src/state/ module #946 case but not a wholesale rename. Proven against a module path that doesn't exist.

📝 Three doc corrections, one to my own text

ADR-011 listed api/globals.sh running set -euo pipefail as a load-bearing ordering constraint. It isn't. The entrypoint sets the same options on its line 2, and system/dependencies.sh repeats them at module #2 — both before api/ (#4) loads. The line earns its place for anything sourcing that file alone, but moving api/ would not turn strict mode off. Struck through rather than deleted, correction dated.

ADR-011 also called "main/ sourced last" load-bearing. True as a description — sourcing main/ first was tried and the suite still passed, because every call resolves at runtime. Reworded.

architecture-map.md and src/api/index.sh both attributed assert_once to api/bashunit.sh. It lives in src/assert/once.sh.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1654 sequential / 1613 parallel-simple-strict — baseline + the 2 new contracts.

[Chemaclass]

🤔 Background

state_test.sh has banned state/ from calling console_results, console_header, parallel and runner since #868/#862. It never banned assert — and assert is now the direction the cycle runs in.

assert -> state is the expected, dominant edge: assertions record their own results. The reverse arrived with assert_once (#921): its marker shares state's per-test reset lifecycle, so state/context.sh calls assert::once_reset and state/counters.sh calls assert::once_is_absorbing. console/test_line.sh reaches into the same feature four more times.

That's a genuine two-node call-graph cycle — invisible to a walk of the source graph, because bash resolves calls at runtime. A full audit of that graph found nothing here.

💡 Changes

• The contract now allowlists exactly those two functions. The exception is deliberate and shipped; anything beyond it is the shape the file already bans. Mutation-tested: an added bashunit::assert::fail_with in state/ turns it red.
• A second test guards the greps themselves.refactor(state): split src/state.sh into a src/state/ module #946 lost a contract this way once: an unmatched glob is handed to grep as a literal, fails No such file, is swallowed by the || true these greps need, and reports a clean pass having scanned nothing. Globbing the module fixed the refactor(state): split src/state.sh into a src/state/ module #946 case but not a wholesale rename. Proven against a module path that doesn't exist.

📝 Three doc corrections, one to my own text

ADR-011 listed api/globals.sh running set -euo pipefail as a load-bearing ordering constraint. It isn't. The entrypoint sets the same options on its line 2, and system/dependencies.sh repeats them at module #2 — both before api/ (#4) loads. The line earns its place for anything sourcing that file alone, but moving api/ would not turn strict mode off. Struck through rather than deleted, correction dated.

ADR-011 also called "main/ sourced last" load-bearing. True as a description — sourcing main/ first was tried and the suite still passed, because every call resolves at runtime. Reworded.

architecture-map.md and src/api/index.sh both attributed assert_once to api/bashunit.sh. It lives in src/assert/once.sh.

✅ Verification

make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1654 sequential / 1613 parallel-simple-strict — baseline + the 2 new contracts.