docs: retain reload drain investigation pending failure logs

quest/m0/test-flakes-2/reload-close-drain.md

@@ -27,4 +27,68 @@ CPU contention and `--rerun-each`, as the epic's earlier JS fixes did
 reproduce in a reasonable number of runs, report that and the runs tried
 rather than changing the test blind.
 
+Decided 2026-11-05 (maintainer, /quest-plan): keep this quest and investigate
+the next captured failure with logs. Abandoning it was rejected: the agent on
+[moq-dev/moq#7764](https://github.com/moq-dev/moq/pull/7764) saw
+`closing the loop keeps the tokens until the session drains` fail once in a
+full-suite run on 2026-11-05. It reported `expect(fin.closed).toBe(false)`
+failing after a 999 ms clock advance, then three passing isolated runs.
+This sighting shows the flake still occurs despite the passing reruns below;
+it does not establish the cause. Keep this wait in the quest so the
+evidence-only [moq-dev/moq#7933](https://github.com/moq-dev/moq/pull/7933)
+can land without a speculative test or production change.
+
+[moq-dev/moq#7764](https://github.com/moq-dev/moq/pull/7764) also disables
+the test (`test.skip`, "load-sensitive fin.closed assertion; investigate drain
+scheduling before restoring") under the repo's disable-a-flaky-test policy, and
+says no follow-up quest was created; this quest is that follow-up. Once that
+skip lands, restoring the test is part of this quest: a skipped test cannot
+meet the Goal or Done condition, and failures can only be captured by
+re-enabling it (for example, with per-dial diagnostics) in a multi-file run.
+
+On the next captured failure, check the dial count and which client closed
+first. In `closing the loop keeps the tokens until the session drains`, the
+fixture overwrites shared `fin.hold` on every dial and lets any
+client's close set shared `fin.closed`, but reads the track from `dials[0]`.
+An extra dial could therefore hold FINs on another client or report another
+client's close. The `teardown by …` fixture does not share this hazard: it
+sets `fin.hold` and `fin.closed` only on the first dial
+(`if (dials.length === 0)`). This is an unverified hypothesis, not an established cause;
+check it alongside setup-group completion before choosing a fix.
+
 Public API: none. Wire: none.
+
+### Reproduction attempts
+
+2026-11-04, recorded in [moq-dev/moq#7933](https://github.com/moq-dev/moq/pull/7933):
+no reproduction on the unchanged claim based on `origin/main`.
+
+- `quest gate bun test --cwd js/net --rerun-each 20 --only-failures`
+  finished with 58,645 passes, 40 skips, and 95 failures across 121 files
+  (476.88 seconds). All 20 reload-file rounds passed. The package
+  batch also exposes unrelated rerun failures in the benchmark import test
+  and connection fallback tests; it is not a clean package validation.
+- A four-file batch (`connection/reload`, `connection/pool`, `connection/close`,
+  and `lite/connection`) filtered to `closing the loop|teardown by|a URL change`
+  with `--rerun-each 5` passed 40 tests, zero failures. A scratch preload
+  stubbed `performance.now()` to advance 5 ms on each budget clock read,
+  forcing the shared stream budgets to yield at each clock check. This
+  stresses budget scheduling only: it neither advances the test's manual
+  clock nor forces an extra dial or another client's early close, so it
+  does not rule out those paths.
+- The earlier lite fixture fix suggests checking setup-group completion,
+  but neither run supports changing this fixture yet. No test or production
+  code changed, per the reproduce-first requirement.
+
+## Waiting
+
+- moq-dev/moq: capture the next reload drain failure with logs before choosing
+  a fix. If #7764's skip has landed, no full-suite failure can occur; restore
+  the test with per-dial diagnostics and reproduce under multi-file load instead. Check full-suite and merge-queue failure reports for the test names
+  above; retain the full command, commit, and assertion/stack output in the
+  report, plus any available session logs. The fixture currently emits no
+  dial/close diagnostics; those logs are not required to resume investigation.
+  Once captured, link that report here, remove this waiting item, and reproduce
+  the failure under multi-file load, capturing dial count and per-client close
+  state to test the hypothesis above. Passing isolated reruns alone do not
+  clear this condition.