Local test suite: unexplained non-determinism (the Synology Drive diagnosis is refuted)

[guycorbaz]

The measured symptom

During stories 4.15 and 4.17 (2026-07-25) the local suite was observed non-deterministic:

• 8 test failures across 5 runs, with identical sha256 on the corpus artefacts and a
  clean git status at every run;
• then 15+ consecutive green runs with no change to the tree.

CI has never reproduced it: 47 runs since 2026-07-22, all green, on clean checkouts.

The recorded cause is REFUTED

Story 4.17's Change Log and the Epic 4 report both recorded the cause as "the corpus was
transiently replaced by a stale server state; cloud-drive-daemon active on the tree."
That cause was never verified. It is false:

Check	Result
cloud-drive-daemon process	none running
Mount type of the tree	local ext4, /dev/nvme0n1p2 — not a network mount
Rows in Synology Drive's history.sqlite touching opencmdb	0 — none, ever
Last Drive sync activity on /home/gcorbaz/Synology/devel	2026-07-12 13:37
First commit of opencmdb	2026-07-17 19:36

Synchronisation of devel was deliberately stopped by the maintainer because it was causing
problems — five days before this repository existed. The tree has never been synced while it
existed. The directory being named Synology/ made the wrong answer plausible on every axis
except the one that was checkable.

What is actually open

The cause of the 8 failures is unknown. Investigation stopped because the wrong explanation
was believed. Candidate directions, none of them checked yet:

• cargo's mtime-based staleness detection on a tree written by several processes;
• concurrent cargo test invocations racing on a shared scratch directory;
• a corpus test reading files while another run rewrites them.

Acceptance

• The symptom is reproduced (or a bounded attempt to reproduce it is recorded as inconclusive).
• The real cause is named with the check that would fail if it were wrong — not a plausible story.
• If it is environmental and not fixable, it is recorded as a named, measured limitation.

Context

• Refuted during Epic 4's retrospective: _bmad-output/implementation-artifacts/epic-4-retro-2026-07-26.md
• The merged story files (4.15, 4.17) keep their period text — history is not rewritten. The
  correction lives in the retrospective, in docs/project-context.md, and here.
• Standing lesson this violated: name the test or the measurement behind every claim — until now
  applied to code only, never to an environment diagnosis.