Phase 39 Gate C Report — Split and Serial (TFP-214, TFP-241, TFP-222, TFP-155, TFP-156)
Date: 2026-10-21
Evidence: benchmarks/2026-10-21-gate-c/. analysis.md has every number below.
Targets: the record’s Gate C table (phase-39-run-concurrency-decision-record.md, “Gate C”).
Setup:
- Variants:
- A (
4a85f1f), before TFP-212;
- C (
2be975c), Gate B’s code;
- C1 (
3b1210c), with TFP-241;
- C2 (
d3e197e), with TFP-214;
- F (
7599789), with TFP-222, TFP-155 and TFP-156, run as the app runs it (units, the serial group second);
- FL, F with the axe scan last.
- One session, interleaved: 04:45 to 10:00, three runs each, one worker, the local Toolshop copies (with-bugs,
then clean), every variant generated from one inspection per copy. Trimmed to the runs the comparisons need.
- A quiet machine: VS Code, Docker and Claude only (Edge and WhatsApp ended by Emil at 04:48, before the first
run). Plane stopped for the session and restarted at 10:00.
For Emil: two decisions, below.
Bottom line
| Comparison |
Judges |
Target |
with-bugs |
clean |
| C against C1 |
TFP-241’s cost |
under 5% |
2.5% slower: kept |
(not run) |
| C1 against C2 |
TFP-214 |
at least 10% faster |
14.9%: met |
20.5%: met |
| A against C2 |
accessibility |
at least 30% faster |
12.1%: not met |
35.0%: met |
| C2 against F |
TFP-222, 155, 156 |
same results; no slower beyond 5% |
Standard 1.6%, accessibility 1.6% faster; same: met |
1.7%, 0.0% faster; same: met |
| F against FL |
where the axe scan runs |
identical axe findings and faster |
identical, 15.6% faster |
identical, 21.1% faster |
- Decision 7 holds in every comparison: the same failing tests in every run of every variant, on both copies, and
0 differences in failures or findings between the variants of any comparison, among the tests that give the same
in every run.
- Scan last meets its rule on both copies. Under Emil’s ruling it becomes the default. That is a code change (the
switch’s default) and a template baseline move, not yet made: for Emil.
- Accessibility misses on with-bugs only, as at Gate B, for the cause Gate B measured. For Emil, below.
| Copy |
Suite |
A |
C |
C1 |
C2 |
F |
FL |
| with-bugs |
Standard |
- |
847 (836, 851, 847) |
868 (861, 868, 874) |
739 (738, 748, 739) |
727 (727, 727, 723) |
- |
| with-bugs |
Accessibility |
207 (205, 210, 207) |
- |
- |
182 (182, 182, 181) |
179 (181, 177, 179) |
151 (157, 151, 150) |
| clean |
Standard |
- |
- |
521 (521, 512, 531) |
414 (416, 412, 414) |
407 (401, 407, 420) |
- |
| clean |
Accessibility |
197 (197, 214, 195) |
- |
- |
128 (133, 125, 128) |
128 (128, 135, 127) |
101 (101, 127, 100) |
- Against Gate B: C’s with-bugs Standard took 847 s (Gate B: 854 s), so the machine was in about the same state.
On with-bugs Standard, A to F is not measured in one session (A has no Standard run here); C to F is 14.2% faster.
- Page loads, with-bugs Standard, run 1: C 319, C1 319, C2 261, F 260. Accessibility: A 80, C2 61, F 61, FL 44.
- F’s split: with-bugs Standard ran as 19 unit files in 2 passes (5 tests in the serial group); clean Standard as
17; accessibility as 20, in one pass (no serial group).
- The spread: at most 4% between a variant’s runs, except clean accessibility, where A spread 195 to 214 s and FL
100 to 127 s.
Decision 7: the same results (analysis.md)
- Failures, every run of every variant: with-bugs Standard 84, accessibility 41; clean Standard 46,
accessibility 19.
- Between variants: 0 tests failing in one variant only, 0 findings in one variant only, in every comparison.
- Tests that vary from run to run of one variant (compared apart, as at Gate B): 5 on with-bugs Standard, the
same kinds as Gate B’s:
- link health’s linked routes (its cutoff);
- the cross-page component check and the pages’ card consistency (bug 4’s random alignment);
- a lab CLS measurement.
None on clean, and none in accessibility.
- The scans, F against FL: every axe scan’s status, findings and manual-review items identical in every run, on
both copies.
The with-bugs / sort message (Gate B ruling 4)
- The check fails in every run of every variant (C, C1, C2, F).
- Its message varies within one variant: C names “Name (Z - A)” in runs 1 and 2, not in run 3. C1, C2 and F name it
in all three.
- So the message depends on the with-bugs app’s state at the moment of the check, not on the code: the same code
gives both messages. This is the reading Gate B recorded, now measured.
For Emil
- Accessibility on with-bugs: 12.1% against 30%.
- The cause is Gate B’s: the axe scan fails on all 20 with-bugs pages. Each failure restarts the worker, so the
checks after it load the page again. On clean, where it fails on fewer pages, the target is met (35.0%).
- With the scan last (FL), with-bugs reaches 27.1% (151 s against A’s 207 s), and clean 48.7%.
- Options:
- a. Accept the gate, with the with-bugs miss recorded. Standard is past its targets on both copies, clean
accessibility is past its target, and with-bugs accessibility is 12% faster with the same results.
Recommended.
- b. Make scan-last the default (decision 2) and judge with-bugs accessibility on FL: 27.1%, still short of
30%.
- c. Apply the rule as locked: not met, reverted. That would give up TFP-214’s Standard gains (15% and 20%)
for a target that one copy’s bug pattern decides.
- The axe scan last becomes the default, as Emil’s rule says (identical findings on both copies, and faster:
15.6% and 21.1%).
- This needs the switch’s default flipped,
check:template-unchanged’s baseline moved in its own commit, and the
template’s accessibility suites changed in order only. Not done without Emil’s go-ahead.