pnpm test glob was directory-enumerated and depended on authors remembering to widen it: latent, no committed test was ever skipped
package.json test script is "node --test src/lib/*.test.ts src/lib/email/*.test.ts" - DIRECTORY-ENUMERATED, not recursive. Any test file in a src/lib subdirectory other than email/ exists, passes locally, and is NEVER RUN by the gate. Found by coder-terra-cc while writing TERRA-24's src/lib/supabase/session-cookie-core tests: 9/9 green standalone, zero of them reached by pnpm test. THE DISCIPLINE THAT MADE THIS A SEPARATE WI IS THE POINT: coder fixed the glob, then REVERTED it before pushing TERRA-22, because package.json is auto-included in every bump commit and the fix would have ridden along silently inside a scope audit had approved. That is exactly right and is the behaviour to preserve - a correct fix smuggled inside an approved scope is still unreviewed. SCOPE OF THE ASK IS WIDER THAN THE ONE-LINE FIX. Do not just make the glob recursive: 1. Enumerate EVERY *.test.ts under src/ and report which are currently reached by the gate and which are not. The count of silently-unrun test files is the finding; the glob is the remedy. 2. For each unreached file, say whether it PASSES when actually run. A test file that was never in the gate may have been silently failing for its whole life, and 'make the glob recursive' would then turn a green pipeline red on unrelated work. 3. Say when each unreached file was added, so we know how long the gate has been narrower than it looks. A green suite that does not run a test file is the same class as a green assertion that can never fail - see the (e-0) finding on TERRA-22, where an extraction that silently matched nothing would have produced a green test measuring a regex built from the string 'undefined'. Both are checks whose passing carries no information, and neither is visible from the pass. Class-S, its own commit, no other change riding along.
Questions
Activity
-
Dispatched to coder-terra-cc as head of the P2 queue while TERRA-22 sits blocked on eq 32 (live session-expiry test with Elazar). Ordered before TERRA-24 because TERRA-24's own test file is one of the files the glob does not reach - fixing the measurement instrument before trusting what it measures.
-
Shipped c1acbba, live 0.18.1 read back cache-busted (dpl_AgcFQj2v3DFTc7aWJQ4gU8NRJBqn, dbOk true). Own commit, package.json only - the smuggling channel stayed shut. THE CENSUS INVERTED THE RISK I FILED THIS AGAINST, and that is the finding. 9 *.test.ts under src/, old pattern reached 8, and the ONE unreached file was coder-terra-cc's own uncommitted TERRA-24 work. ZERO committed test files were ever silently skipped. Nothing had been failing unseen, so the red-on-an-innocent-push hazard I asked them to check for did not exist. The defect was LATENT, not historical. Widening the ask was still right and for a reason the outcome does not show: a clean census and an unrun census are the same green, and only the census distinguishes them. Had one committed file been failing since July, the recursive glob would have broken someone else's unrelated push. IT HAD ALREADY FIRED ONCE AND WAS PATCHED BY HAND: 15eb713 appended src/lib/email/*.test.ts in the SAME commit that added the first email test. The gate has always depended on the author remembering to widen it, and this was the second time it came due. A gate whose coverage is maintained by memory has no coverage property at all - that is what got fixed, not one glob. Fix verified rather than assumed: 107 tests vs 98, and a probe file at src/ root confirmed a top-level src/*.test.ts is now reached. PM error recorded: I set this inProgress and dispatched it ~10 minutes AFTER it shipped, then stated the stale status as fact in the dispatch. coder-terra-cc contradicted it, which is the behaviour I had asked for in the same message. Reading a tracker is not reading the world.
-
pnpm test glob was directory-enumerated and depended on authors remembering to widen it: latent, no committed test was ever skipped
-
TITLE CORRECTED post-close, on coder-terra-cc's flag. Old title: 'test files outside src/lib and src/lib/email exist, pass locally, and never run in the gate' - plural, present tense, reading as a standing condition. The census found exactly ONE such file, uncommitted work of the coder's own, and ZERO committed tests ever skipped. A later reader would have taken the old title as evidence terra shipped untested code. It did not. SECOND TIME IN ONE SESSION for me, after TERRA-37: I closed a WI leaving a wrong claim in the TITLE while the body held the correct finding. A closed WI is read by its title - the body is where someone looks only after the title has told them what happened, so a wrong title on a closed item outlives a wrong body on an open one. The filing hypothesis lives in the title and the close is the moment to overwrite it; I keep treating the close as a body-writing step. The distinction the new title carries: the gate was never narrower than its CONTENTS, it was narrower than its FUTURE contents. Latent defect, not a live one.