Flaky merge-gating test: missionInbox 'reference-stable snapshot' fails intermittently on main (fails Desktop Fast + NuncioCrew Gate)

[devin-ai-integration]

Problem

src/features/home/lib/missionInbox.test.mjs:654 — same inputs return a reference-stable snapshot — fails intermittently, taking the whole Desktop Fast lane and the downstream NuncioCrew Gate with it:

AssertionError [ERR_ASSERTION]: Values have same structure but are not reference-equal
  at src/features/home/lib/missionInbox.test.mjs:665

Evidence that it is flaky, not branch-specific

Measured with the repo's harness loader and a narrowed name filter:

pnpm --filter buzz exec node --import ./test-loader.mjs --experimental-strip-types \
  --test --test-name-pattern='same inputs return a reference-stable snapshot' \
  src/features/home/lib/missionInbox.test.mjs

Branch	Pass	Fail
clean origin/main @ 35af74019	8	1
an unrelated PR branch (#128)	7	2

Identical assertion output on both. origin/main CI run 31362178966 (Desktop Fast job 93373095521) passed; PR #128 run 31395279532 (job 93476664430) failed on it with a documentation-only delta and an empty git diff origin/main -- desktop/src/features/home/lib.

So it fails on main too, at roughly the same rate. It was found while verifying #121 (PR #128) and is unrelated to that work.

Why it matters more than a normal flake

It fails a merge-gating lane (Desktop Fast → NuncioCrew Gate), unlike the known-broken smoke lanes (#109, #110) that reviewers already discount. Every unrelated PR pays a coin-flip tax and, worse, teaches reviewers to wave through a red gate.

What to solve

Find why the memoized snapshot is not reference-stable across identical inputs — a cache keyed on something not fully deterministic (iteration/insertion order, a time or id input, a shared mutable structure being rebuilt) is the usual cause. Fix the identity/memoization, not the assertion: the test is asserting a real invariant, and weakening or skipping it would hide the actual instability.

Definition of Done

• Root cause identified and stated (why the snapshot loses reference stability).
• The instability is fixed in the source, not by relaxing or skipping the test.
• The named test passes in a repeated run (e.g. 20+ consecutive) on the fix branch.
• The test is not weakened, and no new allowance/retry is added to hide it.

[devin-ai-integration]

Fixed in #138.

Root cause: deriveMissionInboxSections folded now = input.now ?? Date.now() into its memo key, so two calls with identical inputs sampled the wall clock independently and missed the cache whenever they straddled a millisecond tick. A 2000-iteration probe on unmodified main showed 8 non-reference-equal results, 8/8 of them coinciding with a differing Date.now() sample, and all 1992 equal ones with an identical sample.

The key is now a pure function of caller-provided inputs (an explicit input.now stays in the key; the implicit clock fallback does not). Narrowed named test: 37/40 pass on unmodified origin/main @ 35af740 vs 50/50 on the fix branch. The test itself is untouched — no skip, retry, or tolerance.