Skip to content

Instantly share code, notes, and snippets.

View ndelangen's full-sized avatar
🇳🇱
yes?

Norbert de Langen ndelangen

🇳🇱
yes?
  • Chroma Software
  • Netherlands
View GitHub Profile
@ndelangen
ndelangen / ci-1443-table-decision.md
Last active October 22, 2026 15:50
CI #1443 table settings menu recovery, proof and delivery

Table settings menu readiness under stalled frames

Tracked class ndelangen/dunezone#1443 PR ndelangen/dunezone#2451 Merge cbea41b3595eb82f23d9b32cbad4bd2026b16b14

Reproduction and repair

Settings Keep Labels On clicks Game menu and then uses a plain Testing Library lookup for Stack counts. Mantine mounts the menu on an animation frame. Two native probes holding frames at that click fail with the missing menuitemcheckbox after the unchanged ten-second wait. The click completes once and there are no page errors. The otherwise identical unheld control passes the full story.

@ndelangen
ndelangen / ci-2411-decision.md
Last active October 22, 2026 15:03
CI #2411 metadata read recovery, proof and delivery status

Published component metadata recovery

Issue ndelangen/dunezone#2411 PR ndelangen/dunezone#2445 Merge 36d04710bc0e44f2cfdc1acd6b3d1a3452a01352

Diagnosis and scope

A transient read of component publication metadata can escape as an image HTTP 503 and Chromium EncodingError. A native workerd/R2 and full Chromium probe reproduces this with the real handler and client and one browser image request. The baseline fails at zero dimensions; the repair returns 200 and decodes the same immutable JPEG at 600 by 600.

@ndelangen
ndelangen / ci-2091-decision.md
Created October 22, 2026 13:50
CI #2091 directory query cold-start repair and deployment proof

Directory-query cold-start timeout, issue #2091

Issue: ndelangen/dunezone#2091 Repair: ndelangen/dunezone#2441 CI map: ndelangen/dunezone#1045

Evidence and cause

The retained WebGPU incident on PR #2180, head a9d73aa1ea798c39db230e51cb8252f544327dd6, ran as Actions run 38463180554, job 115523326765. Native request 6596f4b8c1237c73 for playGames:getGameByAddress used 3.071 seconds of user execution, hit the three-second limit with zero database reads, and threw into the route boundary after both browsers had admitted sockets. The original screenshot shows the error page. Artifact 11848365555 was preserved before any rerun, and its downloaded ZIP matched SHA-256 6af4ef35d85e7d13009284eb210ddcd5f1f69962943300f03bbbf3850a67367a.

@ndelangen
ndelangen / ci-2437-decision.md
Created October 22, 2026 12:32
CI #2437: cold name-query repair, reproduction and deployment proof

Game-name availability cold initialization

Issue #2437, PR #2438, CI map #1045.

The first WebGPU attempt on FAQ PR #2435 failed in playGames:nameAvailability before any database read. The native timing record reached the three-second user execution limit. The create-with-name action itself returned the profanity refusal, but the availability subscription threw during render and replaced Create with the route error boundary. The FAQ diff did not touch backend or shared sources.

The availability query imported the game mutation namespace, including allocation, schema and renderer dependencies. A disposable native backend reproduction reduced backend CPU availability while keeping the native three-second deadline unchanged. Deferring mutation triggers alone and memoizing identity alone did not fix the timeout. The final permanent guard fails against t

@ndelangen
ndelangen / ci-1443-bidding-decision.md
Created October 22, 2026 10:47
CI bidding menu frame-delivery repair, #1443

The bidding menu no longer waits for a native frame

Part of #1443, CI map #1045. PR #2434 repairs the remaining Untimed bidding story exposure.

The actual story clicked Game menu and used plain findByRole for the Timed bidding checkbox. Mantine mounts the dropdown from animation-frame callbacks. If Chromium stopped delivering frames at that click, the menu remained absent until the unchanged ten-second assertion failed. The click itself completed before the failing lookup.

The story now uses the existing waitForFrame mechanism at that lookup. A co-located helper makes the exact sequence available to the permanent real-Mantine regression. Game behavior, checkbox state, bidding timer assertions, focus, transitions, deadlines and zero retries are unchanged.

The final native comparison changes only the lookup in the browser's served helper module. Restoring the old lookup reproduces the missing checkb

@ndelangen
ndelangen / ci-2429-decision.md
Created October 22, 2026 09:32
CI draft/full cancellation boundary, #2429

CI draft/full event ordering, #2429

A stale draft event could cancel a ready PR's full verification because both profiles shared one concurrency group. The original failure occurred on the same commit, with the ready run created one second before the stale draft run.

PR #2432 gives each PR separate draft and full groups. Superseded work still cancels within its own profile. The required full ci_ok gate, cheaper draft checks, deadlines and zero test retries remain unchanged. An existing draft run can finish alongside full CI after readiness.

The paired replay evaluates the actual workflow YAML with GitHub's official expression engine. The original workflow fails two race cases. The repair passes all 14 checks, covering both event orders with full CI running or pending, same-profile cancellation, PR isolation, merge-group fallback and execution of the required gate. Disabling active cancellation still loses pending full work; adding a commit SHA still fa

@ndelangen
ndelangen / timed-bidding-decision.md
Created October 13, 2026 14:38
Accepted optional timed bidding decision and delivery evidence

Optional timed bidding

Accepted by Norbert on 13 October 2026, tracked in #2136 under the Play follow-up map.

Keep the existing bidder. Timing is shared by the table, off by default and controlled from the gear menu. Seated players can change the setting; spectators can observe it.

A smooth ring drains for 20 seconds around the existing bid button, followed by five flashes over five seconds. The server passes after the full 25 seconds. Each new auction waits for its first bid or pass so players can make deals. Enabling timing during an untouched auction preserves that wait. Disabling timing removes its UI and deadline; re-enabling a started auction gives the active faction fresh time.

Delivered by PR #2141 and PR #2151. Required CI, independent spec and standards reviews, and the final CodeRabbit re

@ndelangen
ndelangen / delivery.md
Created October 7, 2026 19:37
Deployed game-name moderation, exact revision and read-only production proof

Game-name moderation delivery

Custom game names block profanity and survive unavailable checks merged through the protected branch with current-head checks and no bypass. Reviewed head is d21f028676f232b2150e137e8ac7949813f239a3; merge commit is 12057413587d38312e77d82f74b7c0ffc33bb2c0.

Current-head CI passed every job, including the regular, catalogue, protocol and WebGPU Play journeys. The creation journey checks stored generated names and final addresses. Independent Standards and Spec reviewers confirmed all findings fixed and no regression in the final follow-up. CodeScene passes all quality gates. CodeRabbit found no actionable comments on the implementation; its final follow-up was quota-limited and covered by both independent reviewers.

Deployment passed the release gate, production deployment and shared developm

@ndelangen
ndelangen / authenticated-browser-proof.json
Created October 7, 2026 19:00
Game-name moderation implementation, local runtime proof and review dispositions
{
"origin": "http://127.0.0.1:35188",
"authenticated": true,
"customAction": {
"gameId": "n9747k0qb2es4wsygjmvdsjgyx8fvcwh",
"moderation": "check_unavailable",
"name": "Dune lantern proof 👍",
"ok": true,
"slug": "dune-lantern-proof"
},
@ndelangen
ndelangen / budget-600-approval.md
Created October 7, 2026 17:59
Fresh broader Jev game-name moderation evaluation with frozen policy and multilingual diagnostics

Norbert explicitly approved increasing the cumulative provider request cap from 400 to 600 in this chat. Preserve the $1 total spend limit and the previously agreed accuracy targets. Existing credits only; no purchase or refill change is authorized.

Continue the protected ledger at 392 attempts and calculated $0.008727432. Freeze the fuller policy with the English-only restriction removed at threshold 0.80. Prepare and freeze 100 fresh English held-out names under the same five group targets, plus 60 diagnostic French, Chinese and emoji cases. Classify the broader cases separately and report their limits; the earlier English-only corpus provides no multilingual accuracy evidence. Count all attempts and retries within the cumulative 600/$1 limits. A failed or incomplete acceptance run returns with evidence.

Norbert's earlier "proceed" authorizes implementing the agreed local checks and backend behavior while the evaluation runs. The provider acceptance decision stays open until measured results are reconcile