|
| 1 | +# Regrade on Solid 2.0.0-rc.21 (2026-11-05, not a round) |
| 2 | + |
| 3 | +The rc.21 repin (41a6b1f) carried the fixes for every Solid issue the eval had filed but did not rerun anything. This |
| 4 | +regrades what the repository holds, with rc.21 from npm and `rxc` from this commit. Round 10's and round 11's desks are |
| 5 | +not in the repository (their run directories stayed in the session that ran them), so they are not here. |
| 6 | + |
| 7 | +## References on the desk9 brief (`harness/grade.sh`: hidden, overlap, 20 fuzz seeds) |
| 8 | + |
| 9 | +| app | hidden | overlap | fuzz seeds with a hole | |
| 10 | +| --- | --- | --- | --- | |
| 11 | +| chain4 .rx (count read with `latest issues`) | 77/77 | 38/38 | 3 | |
| 12 | +| chain4 .rx (count read under the list's region) | 77/77 | 38/38 | 2 | |
| 13 | +| chain4 React | 77/77 | 38/38 | 0 | |
| 14 | +| recipe React (TanStack's optimistic-updates guide) | 77/77 | 22/38 | 7 | |
| 15 | + |
| 16 | +The recipe React desk scores what round 9 recorded for it (22/38), so this setup grades as the rounds did. The chain4 |
| 17 | +.rx reference passes every hidden and overlap test, including "retry after not found" (#4425, #4627), which failed in |
| 18 | +every .rx run from round 5 on. |
| 19 | + |
| 20 | +The `latest issues` reference's three fuzz holes, replayed: |
| 21 | + |
| 22 | +| seed | hole | cause | owner | |
| 23 | +| --- | --- | --- | --- | |
| 24 | +| 1 | the tab count shows 3 beside 2 rows, a close in flight | `latest issues` answers the re-asked truth, not the guess, once a list refresh lands under a pending close | Solid: `latest` should keep a shown guess until its hold commits (solidjs/solid#4689, ruled in #4695 decision 3, not built; DV-19) | |
| 25 | +| 11 | an accepted rename's row shows the old title until the refresh lands | the row was not in the list when the rename was sent (`findIndex` -1): nothing was guessed on it | the app; `guide/async.md` §4 now says so | |
| 26 | +| 15 | an accepted rename's detail heading shows the old title until the refresh lands | the detail was still loading when the rename was sent: nothing was guessed on it (the row guess held) | the app; the same | |
| 27 | + |
| 28 | +Seed 1 was the reference as the checker suggested it: CN-3e refuses the count effect's read of `issues` outside the |
| 29 | +`loading on(filter)` region and suggests `latest issues`. Four references (chain2, chain3, chain4, count) had the |
| 30 | +outside read, predating the rule. By the spec (and Solid's own ruling, #4695 decision 3) `latest issues` answers the |
| 31 | +guess the list shows, so that program is right and the hole is Solid's (DV-19, #4689, not built). The references |
| 32 | +read the count where the list is read instead, in a small component under the list's region: it holds and reveals |
| 33 | +with the rows, it is equally right by the spec, it does not depend on #4689, and it is what 10 of 12 agent desks in |
| 34 | +rounds 10 and 11 wrote. With it, chain4 has two fuzz holes, both the app's (seeds 11 and 15), and chain2, chain3 and |
| 35 | +count pass their own tasks' hidden and overlap tests (`graded/`). |
| 36 | + |
| 37 | +Reducing seed 15 turned up a second Solid bug, solidjs/solid#4690: with nothing rendering a store, an effect reading it |
| 38 | +sees a guess revert at the action's end, before the action's refresh lands, once an outside re-ask has landed under it. |
| 39 | +A rendered reader keeps the guess, and so does an effect beside one; an effect over a source nothing renders reaches |
| 40 | +it (DV-20, S-guess-reask-refresh's effect-only variant; `ref/oracle/repros/4690-guess-reask-refresh.mjs`). Fixed on |
| 41 | +`next` (Ryan, 2026-11-05); recorded until the repin. |
| 42 | + |
| 43 | +## Round 4's .rx apps (18, `results/change-2026-10-28/runs`) on rc.21 and this `rxc` |
| 44 | + |
| 45 | +| | runs | |
| 46 | +| --- | --- | |
| 47 | +| pass every hidden and overlap test | 14 | |
| 48 | +| refused by `STALE_WRITE_BACK` (AC-1e, added after round 4) | 3: small close-rx-1, small comments-rx-2, frontier rename-rx-3 | |
| 49 | +| one overlap test failed | 1: small rename-rx-1, a guess on `detail` with `refresh issues` only | |
| 50 | + |
| 51 | +The three refusals are the three the round 4 write-up predicted. The one failure is the shape `guide/async.md` §4 now |
| 52 | +names ("refresh every question you guessed on"). Nothing in round 4's residue is Solid's. |
| 53 | + |
| 54 | +## Reference suites on rc.21 |
| 55 | + |
| 56 | +Interpreter 114 pass; oracle and compiled 114 pass, 6 recorded divergences (DV-19 at four instants, DV-20 at two); |
| 57 | +server render 107 pass, 7 skip (not rerun after the review fixes); types 156 pass; the crate agrees with ref/ on 321/321 programs. |
| 58 | + |
| 59 | +S-store-8 (two actions' guesses on one row; a declined one reverts alone) had failed on the interpreter since the |
| 60 | +commit that added it (edb76be): the interpreter kept one guess per node with one owner, so the newer action's guess took |
| 61 | +the row and the older action's decline found nothing of its own to revert. Guesses are now entries per key and hold |
| 62 | +(`ref/src/runtime.ts`), as H-11 says; the same-key variant added to S-store-8 passes on all three runners. A new answer that corrects a guess (D-34) marks |
| 63 | +the action's entry spent instead of dropping it, so the action's re-ask still holds the corrected guess (D-35): with |
| 64 | +the entry dropped, a closed row came back between the write landing and the re-ask (S-latest-reask's variant, `4 / 4` |
| 65 | +for `2 / 4`; Solid shows `2 / 4`). The interpreter now follows #4695's D-34: a landing answer is compared with the guess (before, |
| 66 | +only one that changed the truth corrected), the correction re-derives the graph, and the screen, the guess's effects |
| 67 | +and `latest` of a shown guess keep it until the hold commits. |
| 68 | + |
| 69 | +## Setup |
| 70 | + |
| 71 | +`eval/package.json` now lists Playwright; `CHROMIUM=/path/to/chromium` drives a system browser in `harness/driver.mjs` |
| 72 | +and `ref/ssr/run.ts`. `graded/`: the grade output per app above, and `round4-<tier>-<run>.txt` per round 4 app. |
0 commit comments