Skip to content

Commit 2cc81e3

Browse files
ryansolidClaude via Cursorcursoragent
committed
test(web): S1 M5 gate — 50k differential results (§4.10), the two pair regressions pinned
50k × 3 seeds vs a plain next recording: 25 cases differ (2 known), finals identical; 20 are the M3b D-37 display-ahead shape (listed for a ruling as a class); 91501 #8991 a next bug closed at M3a; two regressions where S1 breaks a pair next kept — 3289 #39609 (E8 remount, final screen, M3a) and 7 #45702 (E9 latestDerive, M3b) — pinned test.fails. 47 shadow classes in the generator, untraced. Gate: FAIL on the differential. Co-authored-by: Claude via Cursor <[email protected]> Co-authored-by: Cursor <[email protected]>
1 parent 8ae8829 commit 2cc81e3

2 files changed

Lines changed: 125 additions & 0 deletions

File tree

‎documentation/plans/reader-seat-rebuild.md‎

Lines changed: 74 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -659,6 +659,80 @@ Already gone or out of scope: `markVerdictReader`, `provisionalVerdict`,
659659
reveal record); `stagedScreen` (store.ts still calls it — S4, the store
660660
twins).
661661

662+
### 4.10 Results (S1 M5 gate, 2026-10-29) — NOT CLEAN at 50k
663+
664+
Head 8ae8829e8 (shadow build). Suites: signals 393 files green, 965,387
665+
reads / 242,324 passes / 0 classes; solid 2,388 / 1,546 / 0; web 43,729 /
666+
24,261 / 0. Generator 50,000 × seeds 7 / 3289 / 91501, every pair, head
667+
recorded (`ASYNC_GEN_RECORD`) and compared offline against a recording of
668+
the same cases on the plain `next` @ 5e68925cc build (the comparison is
669+
`differential.ts`'s, run by a script because the two recordings were made
670+
in parallel). 2k × 3 had been clean at every milestone; 50k is not:
671+
672+
| seed | reads | passes | classes | known | screen diffs vs `next` |
673+
| ----- | ----------: | ----------: | ------: | -------------------- | ------------------------: |
674+
| 7 | 327,936,991 | 236,113,560 | **19** | #4427 ×16, #4426 ×10 | **9** |
675+
| 3289 | 427,095,467 | 308,276,567 | **13** | #4427 ×25, #4426 ×15 | **10** (incl. #464 known) |
676+
| 91501 | 381,398,391 | 262,188,619 | **15** | #4426 ×13, #4427 ×13 | **6** (incl. #411 known) |
677+
678+
Survey findings line-identical to the base recording except the law
679+
changes below (seed 7: 623 → 624 cases with findings; 3289: 665 → 665, 117
680+
→ 118 signatures; 91501: 652 → 650, 109 → 108).
681+
682+
**The 25 differing cases, replayed on base, M2 (490a387ba), M3a
683+
(b6ea9ceb1), M3b (143d829bf) and the head's prod build** (the shadow does
684+
not alter them — prod and shadow screens identical):
685+
686+
- **Final screens identical to `next` in all 25**; every difference is an
687+
intermediate (mid-hold) screen.
688+
- **20 are one shape, introduced at M3b**: under a display-ahead
689+
`<Show when=latest(x)>` the lane work shows the committed value (or an
690+
inner `<Show when=x()>`'s fallback) where `next` showed the staged one —
691+
e.g. `<span>vs0:1 ls0:1</span>` → `<span>vs0:0 ls0:1</span>`; `[F]` for
692+
`<span>vs0:1</span>`. Screen-identical to base at M2 and M3a. This is the
693+
#411 shape (D-37, "lane work sees the screen") the adoption produced and
694+
§4.6 pinned as a `next` bug; at 50k it is a class of 21, not a case.
695+
**Not ruled as a class** — listed for the ruling: seed 7 #9846 #13595
696+
#19106 (the error surfaces one settle earlier) #20324 #21348 #39486
697+
#42058 #48845; 3289 #3598 #8992 #12559 #22669 #28591 #30977 #34631
698+
#49646; 91501 #3494 #5871 #14705 #45722.
699+
- **2 `next` bugs closed** (laws removed): 3289 #464 (E9, M3b, §4.6);
700+
**91501 #8991** (E9, closed at M3a — `latest()` through a memo showed
701+
`lm0:3` where the leaf showed an empty span under
702+
`<Show when=isPending(s0)><Show when=isPending(m0)><Show when=m0()>`;
703+
S1 shows the empty span in both).
704+
- **2 regressions — S1 breaks a pair `next` kept** (pinned `test.fails`,
705+
`packages/web/test/async-gen/gate-50k-pairs-3289-39609-7-45702.spec.ts`):
706+
- **3289 #39609 (E8, remount) — introduced at M3a.** The reader born in
707+
the flush that applies the write ends with the inner
708+
`<Show when=m0()>` content missing from its **final, quiet** screen
709+
(`<span data-a0=lm0:3></span>` absent); the reader born before shows
710+
it; `next` shows it in both. M3a's changes are the pass side: seat
711+
world as module state, born-held on F-transitive B, `creatorReads`
712+
deleted, `route` folded — B was ruled "no screen moves" on 2k × 3; this
713+
is a final-screen move at 50k.
714+
- **7 #45702 (E9, latestDerive) — introduced at M3b.** `latest()` through
715+
a memo shows the content frame at the settle after `t` where the leaf
716+
form shows nothing; `next` shows nothing in both; finals agree.
717+
718+
**Shadow classes (47)** — all in the generator, none in the suites; not
719+
yet traced to a milestone (a 50k shadow run per milestone is ≈ 95 min
720+
each). Rows: **V6** (a reader of a verdict-lane node: `vlane`, `oldVlane`)
721+
in four variants — slot `V6→L0` / `V6→D0` expected C got Lv (the table
722+
serves the lane's value where the deciders served committed; flags
723+
`shown`, `parentMerged`, `lheld`); slot `V6` expected Lv got Dv; marks
724+
`LST` vs `leave+LST` and `laneRead` vs `leave+laneRead` (the oracle fires
725+
a `leave` the table does not); **L4** expected P got C (L posture, `h ip`
726+
— a held reader under `isPending`) ×3 classes; **B2** born held (old
727+
false, new true; world 2/0/0) ×7 — F-transitive B's known class
728+
(§4.5). `#4427`'s known filter (`parentMerged` identity reset) catches 54
729+
of the V6 family; these are its unmatched variants.
730+
731+
**Gate verdict: FAIL on the 50k differential** (two regressions, one class
732+
of 20 awaiting a ruling, 47 untraced classes); suites, `it.fails`
733+
inventory and `scripts/size` as reported in the PR body; clause (c) is a
734+
pending ruling. Nothing merged.
735+
662736
## 5. Swap plan
663737

664738
Three to five PRs, serial, each in its own worktree (brief §4 item 4):
Lines changed: 51 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,51 @@
1+
/**
2+
* @vitest-environment jsdom
3+
*/
4+
// Two generator cases the S1 M5 gate (50k × 3 seeds, design doc §4.10)
5+
// found where S1 breaks an equivalence pair `next` @ 5e68925cc kept.
6+
// Pinned as `test.fails` until traced: the law is the generator's (E8, E9),
7+
// the fix is not ruled.
8+
//
9+
// seed 3289 #39609 (E8, remount — introduced at M3a, b6ea9ceb1): a reader
10+
// born in the flush that applies the write (`remount`) ends with the inner
11+
// `<Show when=m0()>` content missing from its final, quiet screen; the
12+
// reader born before the write shows it. `next` shows it in both.
13+
// [s0=0 s1=2 m0=async(latest(s0) !%4=2)]
14+
// <span fx[m0()]></span><span>m0()</span><span>m0() s0()</span>
15+
// <Show when=m0() fb><Show when=latest(m0)><Loading><Show when=m0()>
16+
// <span fx[latest(m0)]></span></Show></Loading></Show></Show>
17+
// :: f t S2 S3 S7 A(s1=0) t S3 s0=2
18+
//
19+
// seed 7 #45702 (E9, latestDerive — introduced at M3b, 143d829bf): under
20+
// `<Show when=latest(s0)><Show when=latest(m1) fb>`, `latest()` read through
21+
// a memo shows the content frame (`<span></span><span></span>`) at the
22+
// settle after `t` where the leaf form shows nothing; `next` shows nothing
23+
// in both. Finals agree.
24+
// [s0=0 m0=async(s0 !%4=1) m1=async(s0 !%4=3) m2=async(m0 !%4=3)]
25+
// <Show when=latest(s0)><Show when=latest(m1) fb><span split[m1()]></span>
26+
// <span>latest(m2)</span></Show></Show>
27+
// :: s0=1 s0=0 s0=0 m S7 S0 s0=1 t s0=3
28+
import { expect, test } from "vitest";
29+
import fc from "fast-check";
30+
import { runScenario } from "./run.js";
31+
import { scenarioArb } from "./scenario.js";
32+
33+
test.fails(
34+
"3289 #39609 (E8): a reader remounted in the flush shows the same final screen as one born before it",
35+
async () => {
36+
const sc = fc.sample(scenarioArb(), { seed: 3289, numRuns: 39610 })[39609];
37+
const before = await runScenario(sc, {});
38+
const remount = await runScenario(sc, { remount: true });
39+
expect(remount.final).toBe(before.final);
40+
}
41+
);
42+
43+
test.fails(
44+
"7 #45702 (E9): latest() through a memo and at the leaf show the same screens",
45+
async () => {
46+
const sc = fc.sample(scenarioArb(), { seed: 7, numRuns: 45703 })[45702];
47+
const leaf = await runScenario(sc, {});
48+
const memo = await runScenario(sc, { latestDerive: true });
49+
expect(memo.screens).toEqual(leaf.screens);
50+
}
51+
);

0 commit comments

Comments
 (0)