ekp/postmortem/20260728-test-state-isolation.md

Make Test State Ownership Explicit

Context

The test named as a parshape C-bypass check actually configured first-line indent, which C 1.5 supports. It passed in the alphabetic full suite only because an earlier C-parameter test left ekp-use-c-module disabled. The shared fixture restored a hand-picked penalty subset rather than the state its callers could mutate.

Decision

The fixture dynamically binds every public EKP tunable plus spacing state, so arbitrary assignment inside a test is automatically scoped. Dispatch regressions go through the public formatter and use an observable forbidden C call, rather than asserting only an internal eligibility predicate.

Keep three complementary entry points: normal order for speed, a seeded permutation for cross-test leakage, and one fresh process per test for absolute isolation. The runners use the public ert-select-tests selector contract.

Alternatives Rejected

Reset more variables to hard-coded defaults

That overwrites caller configuration, must be updated whenever defaults change, and repeats the partial-ownership mistake.

Test only ekp--c-available-p

It does not prove that the installed public dispatcher uses the same input shape or filtering logic.

Depend on alphabetical order

Stable order is useful for reproducibility, not as a correctness boundary.

Consequences

• Leaked test assignments are restored to their incoming values.
• Parshape bypass is proven through the same path users call.
• A failed randomized seed can be reproduced with EKP_TEST_SEED.
• Full process isolation remains intentionally slower and opt-in.

Verification

The original isolated C-loaded control failed 0/1. The focused repaired matrix passed 3/3, a non-default permutation passed 108/108, and all 108 tests passed one-by-one in fresh Emacs processes.

Rollback

Restore the old fixture and test, then remove the two runners and CI lane. No production or persisted data needs migration.