flake: Node wheel-permission focus assertion fails in CI #1125

CI2457 fails one existing wheel-permission test at 5f84413f165df4d65694a8ee062af3d7366c6160. The same file passes locally without a source change. Investigate this as a possible timing/platform flake; the local and CI environments differ, so the local pass does not establish a successful unchanged CI retry.

• Failed job: CI2457, Node shard10/16, Ubuntu, Node22.23.4.
• File: packages/cli/tests/repl-wheel-migration.test.ts.
• Case: permission pages from a control-only window to its explanation and final choice.
• Assertion at line84: after eight downward wheel ticks over the focused first choice, the screen must still show > [Choice 00]. It received false.
• The remaining five cases in that file passed in the failed CI invocation. All29 assigned files ran; this file was the shard’s only failure.

Expected: true
Received: false
at packages/cli/tests/repl-wheel-migration.test.ts:84:50

One local run at the identical commit passed all six cases under cached Node22.23.2 on macOS, with a clean source checkout before and after. This is useful contrast, not an identical Linux/Node22.23.4 reproduction:

node node_modules/tsx/dist/cli.mjs --tsconfig tsconfig.node.json --test --test-concurrency=1 packages/cli/tests/repl-wheel-migration.test.ts

# tests 6
# pass 6
# fail 0
# duration_ms 36437.81925

The first complete CI failure log and the local run’s first output are retained. No CI retry, test weakening, source repair, exclusion or timing increase has been applied. The wheel test is unchanged by the #1115 keyboard fix. Establish the triggering cause with a discriminating regression before changing its focus-preservation assertion. A later green run restores delivery health but does not close this issue.

Related #1054 concerns a different Bun History/final-selection case in the same file; this record isolates the Node permission-focus assertion and its evidence.

Additional runner evidence on PR #1038

At unchanged commit 47f0101ed9979e73650a2885b25305d449307564, weight measurement 42139579929, attempt 1, Bun job fails this file under Bun 1.4.0 (34cbb9a40) on ubuntu-latest. All ten Bun shards pass in CI 42139412370, attempt 1 at that same commit. The full first measurement log is preserved (6,917,572 bytes; SHA256 2b7092360f8275ca26496cb65d0c953a01b0077cd9fb5129cebb510fa1961dc9). The original test-weights-bun-attempt-1 artifact, ID 13470241356, records failure after 122 of 440 files; ZIP SHA256 e073aa252bfca78e2a3f21d01b1b46d556e798b8f8015dc14c07cfb11f5fb1e2. This failed record is ineligible for assembly.

Two assertions fail in this invocation:

• permission pages from a control-only window to its explanation and final choice: line 84 expects > [Choice 39] after 39 ArrowDown inputs; received false. This is a different exact assertion from the historical Choice 00 report, so a common cause is not established.
• History alone scrolls above a held question and restores its caret: line 324 expects the prior marker to disappear after a downward wheel input; it remains visible. This differs from the historical final-selection assertion tracked in flake: Bun wheel-history journey intermittently misses the final selected row #1054.

The other four cases pass. Measurement uses full deno task setup, while ordinary CI uses its documented preparation. Preparation and input/render ordering are investigation variables, not a proven cause. No rerun, timeout increase, exclusion or repair has been applied. These findings do not expand the separately approved TODO focus correction. Related: #1125 and #1054. Keep the issues open until causal repair and regression evidence exist.

Fresh integrated-head Bun recurrence

At signed integration head b66436fe34586721e8be8eab2e5f8958fad061fd, CI42199871873 attempt1, Bun4/10 fails repl-wheel-migration.test.ts:84: > [Choice39] is absent after39ArrowDown inputs. This matches the later Choice39 assertion recorded above, not the original Choice00-after-wheel assertion. Runtime Bun1.4.0+34cbb9a40, runner ubuntu-latest (Ubuntu26.04.1).

The complete first log is404154UTF8 bytes, SHA256 0a80961a9bb309162e8bcf0bd46542d0fd1bae5bf9997d9d0cc9804964668f0c, preserved before any retry. All17Node shards succeed at this same head; different runtimes do not isolate a cause. The full CI attempt completed with this and two Deno test failures. No CI rerun, local reproduction or source correction has followed. Keep all focus/navigation/provider-decision assertions, current deadlines and the original source history. This is recurrence evidence, not a causal fix or proof that every historical wheel/History failure shares a cause.

Pre-merge occurrence audit at 5d4f4fb

The original Planner has now audited these occurrences against signed 5d4f4fb97894f34b3f968bee4de827195efa6b74. The current directional walk acknowledges each exact choice and holds one middle presentation as a negative control. That covers the demonstrated acknowledgement hazard; it does not establish the cause of either historical Choice39 failure. The original Choice00-after-wheel assertion and the separate History downward-wheel assertion remain unchanged and unexplained. Wheel routing uses the current scroll map and admitted viewport window; recovery pointer-map evidence does not diagnose it.

Current CI42248271091 attempt1 completed green without a rerun. This changed-head success does not close any historical occurrence. The user has required pre-merge resolution of the overlapping cases. A bounded wheel-target/focus/window capture on matching Linux Node/Bun toolchains is proposed, with first-failure stop and no guessed repair, larger deadline, key substitution, loop or automatic retry. No diagnostic execution or new source allowance has been approved; this issue remains open.

Bounded Linux diagnostic: failures not reproduced

Diagnostic42284168554 attempt1, isolated signed 4d9fbcd4573119912e5436a5b87643dd900e3c1b based on product 5d4f4fb9, completed the Node22.23.4 permission selection (1pass) and Bun1.4.0+34cbb9a40 permission/History selection (2pass) on Ubuntu26.04.1 after one setup. The original assertions and waits were retained. Selected traces show the initial permission wheel clamped at from47 around the focused Choice00 range47..48, and the relevant History wheel moved offset0→3, with the reverse returning3→0. These are expected successful transitions, not captured historical failures.

The original Planner's post-run reconciliation confirms not reproduced, with no causal closure or speculative wheel patch. Production sequence and test ordinal differ; correlations use order, coordinates, frames and window/focus facts, not equal counters. The same invocation later stopped on a separate scratch paging compile error; no diagnostic was rerun. Complete original ZIP/artifact13537495052 (SHA256 82429cad376a826149e04a32e2ce3d8221b8c95a88aa0d84a9dcf5bd7b85c2e2) and full raw logs/traces are retained. No temporary diagnostic code is adopted into PR#1038. Passed wheel selections will not repeat for confidence; this issue stays open with the historical occurrences unexplained.

Human disposition of the remaining historical uncertainty

After the completed occurrence audit and diagnostics for PR#1038, the user explicitly accepted the documented residual wheel/navigation/History uncertainty and deferred this investigation: “Ok, we’ll deal with them later.” The added pre-merge flake-acceptance hold is resolved. This issue remains open, and its distinct historical causes remain unexplained. No additional wheel run, guessed production fix or causal closure is claimed. Current product technical reviews/CI and the separate successful paging control retain their actual scope; no scratch code is adopted. See the retained PR status.