Five distinct specs failed across two runs of identical code with no overlap between the sets, so which test fails is decided by the scheduler. The cost is already being paid: #245 skipped the only end-to-end check that adding an item reaches the database, purely to get main green.
The design separates two mechanisms that had been treated as one. Unchecked lookups into shared collections — `collection.find(...)` dereferenced immediately, found at eight or more sites — are a defect regardless of concurrency: when the row is missing the test dies with "Cannot read properties of undefined" naming test plumbing rather than failing an assertion that says what it wanted. Load-induced timing is the other, and is the larger share of what has actually been observed: three of four local failures and the CI one are assertions in a spec's own browser context that nothing else can touch.
Two claims from earlier in this investigation are retracted in the document rather than quietly dropped. The verification-resend limiter is not a shared axis — it is keyed per customer and every test registers its own — and the suite contains no snapshot-style assertions, so there is nothing to convert to web-first. Both were stated as fact on the issue, and both would have justified work that was not needed.
Per-worker databases are ruled out structurally: every worker talks to one backend on :3000, so isolation there means N backends, not N databases. Worker-count reduction is deliberately deferred rather than taken now, because applying it at the same time would mask whether fixing the defects worked.
The honest limit is recorded too. CI's load cannot be reproduced here on demand, so the timing changes rest on reasoning rather than a red-to-green demonstration, and the success criterion is several consecutive green runs rather than one.
Ref #241
bermudalamb
f8b2b68f0d