docs(handoff): the store-api flake is DIAGNOSED as fsync contention — the doc's next-probe was the refuted thread

[ZacxDev]

Found by reconciling the canonical tmux-webapp handoff against live state at resume time.

1. A conflict marker was committed into the doc

Line 699 carried ||||||| parent of e1d1318f (…) — a diff3 marker left from the #1162
merge, sitting directly above the docker.io correction block. Removed.

2. The store-api flake blocks were superseded and the doc did not know it

grep -ci 'ci-repro|fsync' over the doc returned 0, while three PRs had already merged:

PR	squash	what it found
#1181	0c333846	it is fsync CONTENTION, not seed/ordering — with a dev-host reproducer
#1190	634c328a	the empty read was reported as a SECOND response (12 assertions)
#1193	48a5540e	the hang guard SAMPLED its own arming instead of waiting for it

The doc still read "Leading hypothesis: non-determinism … seed/ordering" and sent the next
reader to open a PR for fix/xdist-parametrize-values-deterministic. That is the refuted
thread. Following it costs a session — which is exactly the failure mode this doc documents
twice about its own open-investigation blocks.

Both blocks are kept verbatim and marked SUPERSEDED: their eliminations are still sound,
it was the framing that was wrong. A new ### ✅ DIAGNOSED block points at
scripts/ci-repro/README.md as the canonical maintained write-up rather than copying it,
and carries the two fixes-that-look-right-and-are-not.

Deliberately not written as "fixed"

Per this doc's own shelf-life rule. Measured 2026-09-01 after all three merged: #1197 is
red on TestAHungRoundTripSAYSWhichSideBlocked::test_a_stall_in_the_FSYNC_region_is_NAMED —
a fourth distinct case — while #1199 passed the same tier. Recorded status: mechanism
identified and reproducible, three contributing defects removed, no consecutive greens in the
failing tier yet.

Two live facts the doc had stale

• ZacxDev/homelab-infra#572 is MERGED (2026-08-31 18:36Z); the doc framed it as open.
  Its own warning is preserved — it fixes one of the two TaskRunTimeout causes.
• Rank 8a's "both clients are stale again by construction" was tested and is wrong.
  Re-measured 16:29Z: server 0.8.21, workbench client 0.8.21, laptop client 0.8.21, and
  the cross-host round trip moved a number (view create on the laptop → view ls on the
  workbench → view rm, [] → 1 → []). Recorded as a dated measurement, with the point that
  "stale by construction" is a hypothesis to test, not a fact to carry.

Gate

Docs-only diff; no code path is reachable from it. The six content/handoff gates a docs
change can red pass in the dev-host tier — 652 passed (test_no_captured_text,
test_no_captured_markup, test_no_public_ips, test_no_client_hostnames,
test_handoff_doc, test_no_real_launchers). The sandbox tier is the one that gates the
merge; it runs on this PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_019kRyYxB956h5XfpwoX2sKX