7299fe80

docs: name which run each reproduction rate came from

a73x 2026-08-08 18:40

Commit message

docs: name which run each reproduction rate came from

Two failure rates two-fold apart sat in the record with nothing saying they
came from different batches, which reads as one number transcribed wrong.
They are three separate reproductions: 21/200 from the investigation, 25/40
from the implementer at smaller N, 3/8 as the review's control — each driven
to zero by the fix, none sharing a harness with the others. The zero-leaks
observation belongs to the 200-run batch specifically, and now says so.

Three independent reproductions is also the reason the section can say
"observed" without qualification, which is worth stating where the numbers
are rather than leaving a reader to notice it.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

docs/decisions.md

Old		New
@@ -1111,13 +1111,21 @@ fit is that each explained half the evidence:
1111	- The transient was `drainPending`. It gives up the first time a wakeup	1111	- The transient was `drainPending`. It gives up the first time a wakeup
1112	retires nothing, which misreads a coalesced or flow-control-only	1112	retires nothing, which misreads a coalesced or flow-control-only
1113	acknowledgement as a finished peer. Reproduced on HEAD under four-way	1113	acknowledgement as a finished peer. Reproduced on HEAD under four-way
1114	contention on one core at 25 failures in 40 runs, always the same	1114	contention pinned to one core, three times independently, always the same
1115	test, and driven to 0 in 40 by the fix. Observed, not inferred.	1115	test: 21 failures in 200 runs (the investigation's batch), 25 in 40
		1116	(the implementer's, at smaller N), and 3 in 8 as the review's control.
		1117	Driven to zero by the fix in every one — 0/200, 0/40, 0/12 respectively.
		1118	Observed, not inferred.
1116	- The leaks were the ngtcp2 abort — but reaching the test suite through	1119	- The leaks were the ngtcp2 abort — but reaching the test suite through
1117	other concurrent processes sharing `/tmp`, not through the run that	1120	other concurrent processes sharing `/tmp`, not through the run that
1118	showed the failure. That is why the two never correlated: the 21 HEAD	1121	showed the failure. That is why the two never correlated: the
1119	failures produced zero leaks, because an assertion failure runs its	1122	investigation's 21 failures produced zero leaks, because an assertion
1120	defers and only a signal kills a process before they can run.	1123	failure runs its defers and only a signal kills a process before they can.
		1124
		1125	Three reproductions at three sample sizes, by three parties who did not
		1126	share a harness, is the reason this section can say "observed" without
		1127	qualification — and naming which batch each number came from is what keeps
		1128	the two 2x-apart rates from looking like one number transcribed wrong.
1121		1129
1122	So the leaked temp directory was a true clue pointing at a real defect, and	1130	So the leaked temp directory was a true clue pointing at a real defect, and
1123	the failure it was found next to had a different cause entirely. Neither	1131	the failure it was found next to had a different cause entirely. Neither
@@ -1178,8 +1186,9 @@ reset by any progress) rather than acting on one, and by bounding the poll
1178	slice with ngtcp2's own next deadline, floored at 1ms.	1186	slice with ngtcp2's own next deadline, floored at 1ms.
1179		1187
1180	**Both halves were handed over as required; measurement says otherwise, and	1188	**Both halves were handed over as required; measurement says otherwise, and
1181	the measurement is what goes in the record.** Four-way contention on one	1189	the measurement is what goes in the record.** This is the implementer's
1182	core, 40 runs each:	1190	40-run batch (the same one quoted above as 25/40), four-way contention on
		1191	one core, run per configuration so the halves could be separated:
1183		1192
1184	\| \| failures \|	1193	\| \| failures \|
1185	\|---\|---\|	1194	\|---\|---\|