Re-look at #34 click-to-focus in light of #55: is the wheel-path stray node click the same hard-coded-coordinate capture?

[mohidmakhdoomi]

Goal

Re-examine the click-to-focus test (tests/e2e/matrix.spec.ts, "click-to-focus fixes the node, animates the camera, and reset restores the view") in light of #55, which proved that a stray node capture at a hard-coded screen coordinate — not browser/rasterizer nondeterminism — was the real mechanism behind a long-standing "flake".

Two questions, in order:

1. Does Firefox e2e flake: "background drag" rotation (matrix.spec.ts:224) — synthetic-input-delivery nondeterminism (survives on hardware) #55's mechanism also explain this test's stray-click recovery loop? (Primary — new information, concrete and testable.)
2. Is the residual defensive scaffolding still earning its keep now that Flaky CI e2e: click-to-focus (matrix.spec.ts:235) intermittently fails on SwiftShader (~50% observed 2026-07-20) #34's hover-first fix has landed and Firefox e2e flake: "background drag" rotation (matrix.spec.ts:224) — synthetic-input-delivery nondeterminism (survives on hardware) #55 has retired the sibling folk diagnosis? (Secondary — cleanup, gated on the answer to 1.)

This issue is not a claim that #34 is unfixed. #34 received a genuine root-cause fix (65a1ad7, 2026-07-20): the library resolves hover only in the render loop's throttled raycast and defers onClick a frame past pointerup, so under software WebGL a bare click could resolve against a not-yet-committed hover; settleHoverThenClick forces a committed hover raycast before pointerdown. That mechanism is documented in tests/e2e/pointer.ts and is not in question here.

Motivation

#55's finding. The Firefox background-drag flake was blamed for years on "synthetic-input-delivery nondeterminism". Instrumentation refuted that (12/12 pointermoves delivered in every failure) and proved H1 stray node capture: the hard-coded drag start (150, 450) intermittently landed on a node whose screen projection had grown after the wheel-zoom, DragControls grabbed it, and the Trackball was disabled. Rasterizer-independent (CPU-side three.js raycast → survived on RTX 3080), per-run-random (random layout seed), Firefox-dominant (zooms ~18–22% closer → bigger projections).

The link to this test. This test contains its own unexplained stray-node problem, and it is structured around a hard-coded screen coordinate in a moving scene — the exact anti-pattern #55 fixed. From matrix.spec.ts:

// Wheel from a corner: while zooming through the rotating cloud, nodes sweep
// under the pointer, and the hover pipeline can otherwise register a stray
// node click that stops rotation and fixes a node prematurely.
await page.mouse.move(60, 540);

and, immediately after:

// The wheel path itself occasionally registers a stray node interaction
// (observed in both engines) that fixes a node and stops rotation. [...]
// on a stray, restart from a fresh page instead of aiming into a corrupted state.

That is a static pointer, moving scene variant of #55's static coordinate, changed scene. The parked corner (60, 540) is asserted-empty by nothing; the loop then wheels the camera 60 bursts closer, growing every node's projection, while the cloud auto-rotates. Nodes sweeping under a parked pointer is precisely how a hover-pipeline stray click would arise. The current handling is a recovery loop (detect fixedNodeCount !== 0, full page reload, retry, up to 2×) — i.e. we detect and recover from the symptom without having proven the cause.

Why now. #55 left behind exactly the tooling to settle this, deliberately retained for future background-point triage:

• nodeOccupancyAtPoint / __graphNodeOccupancyAtPoint in tests/e2e/graph-handle.ts — the DragControls-mirroring hit test that was the H1 discriminator.
• pickBackgroundDragPoint — max-edge-clearance point selection with a pixel-margin floor.
• The out-of-tree amplified-reproduction harness under tests/diagnostics/55-drag/ (own playwright.diag.config.ts), a re-runnable template for this investigation.

Scope

Phase A — Instrument the stray (primary).

• Sample nodeOccupancyAtPoint(60, 540) across the zoomIntoClickRange burst loop and correlate occupancy hits with fixedNodeCount 0→1 transitions, using Firefox e2e flake: "background drag" rotation (matrix.spec.ts:224) — synthetic-input-delivery nondeterminism (survives on hardware) #55's instrument-first method (out-of-tree diagnostic harness; canonical suite untouched during investigation).
• Amplify to a decisive reproduction count of the stray, then confirm or refute: is the stray a raycast hit at the parked corner, or something else?
• Explicitly discriminate hit-what-I-meant (occupancy) from did-input-arrive (wheel/pointer counters), the discrimination that broke Firefox e2e flake: "background drag" rotation (matrix.spec.ts:224) — synthetic-input-delivery nondeterminism (survives on hardware) #55 open.

Phase B — Fix or justify (conditional on A).

• If confirmed: replace the parked corner with a probe-verified clearance point maintained across the zoom (reuse/extend pickBackgroundDragPoint; the point must stay clear as projections grow, so re-probing between bursts may be needed). The recovery-reload loop then becomes dead code and can be retired.
• If refuted: document the actual mechanism and keep the recovery loop, with the refutation recorded so the next reader does not re-open this.

Phase C — Re-qualify the residual scaffolding (secondary, gated on B).
Only after A/B settle, audit each item on evidence — retire what no longer earns its keep, keep and justify the rest:

• test.setTimeout(240_000) — the suite's only 240 s override.
• The 60-burst adaptive zoom loop with 600 ms / 2 500 ms stall pauses.
• The 2× stray-click recovery reloads.
• The 6-attempt aimed-click retry loop.
• CI retries: 2 (playwright.config.ts:132) — landed as Flaky CI e2e: click-to-focus (matrix.spec.ts:235) intermittently fails on SwiftShader (~50% observed 2026-07-20) #34's mitigation before the root-cause fix existed. Measure the current flaky-marked rate in merged Playwright reports first; do not remove it on vibes.

Constraints and invariants

• Do not weaken assertions. MOTION_FLOOR, local retries: 0, and the real synthetic gesture path stay as strict as they are. Firefox e2e flake: "background drag" rotation (matrix.spec.ts:224) — synthetic-input-delivery nondeterminism (survives on hardware) #55 fixed the cause rather than relaxing the test; same bar here.
• .github/workflows/validation.yml byte-for-byte unchanged; the absolute CI → workers: 1 guard stays intact. Any change to CI retries is a config-line change with recorded evidence, nothing more.
• Do not trim qualified waits without a proven mechanism replacing them (standing constraint from CI: Validation workflow takes ~15 min; shard the Playwright e2e suite to cut wall clock to ~5-6 min #30, restated in Flaky CI e2e: click-to-focus (matrix.spec.ts:235) intermittently fails on SwiftShader (~50% observed 2026-07-20) #34). Retiring scaffolding requires evidence that the thing it guarded against cannot happen — not merely that it has not happened lately.
• Heavyweight diagnostics stay out-of-tree under tests/diagnostics/ (the FR2 pattern from Firefox e2e flake: "background drag" rotation (matrix.spec.ts:224) — synthetic-input-delivery nondeterminism (survives on hardware) #55); the committed tests/e2e/ delta should be the fix plus any cheap probe.
• Reproducibility contract: no new dependencies, no lockfile/toolchain movement. npm run validate remains the green gate.
• Test-harness change expected; no app-code change anticipated (as with Firefox e2e flake: "background drag" rotation (matrix.spec.ts:224) — synthetic-input-delivery nondeterminism (survives on hardware) #55, this is a test-assumption defect if it is anything).

Definition of Done

• Stray-node mechanism in zoomIntoClickRange confirmed or refuted with instrumented, verbatim evidence (occupancy ⟺ fixedNodeCount 0→1 correlation table, in Firefox e2e flake: "background drag" rotation (matrix.spec.ts:224) — synthetic-input-delivery nondeterminism (survives on hardware) #55's format).
• If confirmed: parked-corner coordinate replaced with a probe-verified clearance point; recovery-reload loop retired.
• Qualification across both engines × both rendering paths (SwiftShader + native-GPU lane), serial and parallel, at retries: 0 — Firefox e2e flake: "background drag" rotation (matrix.spec.ts:224) — synthetic-input-delivery nondeterminism (survives on hardware) #55's bar.
• Current flaky-marked rate for this test measured from merged Playwright reports since 65a1ad7, and recorded.
• Each retained scaffolding item has a one-line evidence-backed justification; each retired item has its removal justified. No silent trimming.
• npm run validate green on a clean detached-HEAD worktree with real npm ci (lessons-critical clean-checkout rule).
• Flaky CI e2e: click-to-focus (matrix.spec.ts:235) intermittently fails on SwiftShader (~50% observed 2026-07-20) #34 referenced and its record reconciled — including whether CI retries: 2 remains warranted.

Non-goals

• Changing the SwiftShader-vs-hardware gate model ("the lane is never the gate" — Add an opt-in native-GPU local e2e lane (WSL2 Mesa d3d12 / hardware WebGL) #44 Decision 1, Add Firefox hardware WebGL to the native-GPU local e2e lane (WSL2 Mesa d3d12) #52 Decision 1).
• Flipping the global DEFAULT_LOCAL_WORKERS default (blocked on deterministic Chromium SwiftShader parallel contention; see Make the native-GPU e2e lane default to parallel workers (SwiftShader gate stays serial) #56 and Firefox e2e flake: "background drag" rotation (matrix.spec.ts:224) — synthetic-input-delivery nondeterminism (survives on hardware) #55's follow-ups).
• Re-litigating Flaky CI e2e: click-to-focus (matrix.spec.ts:235) intermittently fails on SwiftShader (~50% observed 2026-07-20) #34's hover-first fix, which is proven and documented.

References

• Firefox e2e flake: "background drag" rotation (matrix.spec.ts:224) — synthetic-input-delivery nondeterminism (survives on hardware) #55 — background-drag flake: stray node capture; pickBackgroundDragPoint, nodeOccupancyAtPoint, tests/diagnostics/55-drag/. codev/reviews/55-firefox-background-drag-flake.md.
• Flaky CI e2e: click-to-focus (matrix.spec.ts:235) intermittently fails on SwiftShader (~50% observed 2026-07-20) #34 — original click-to-focus flake; CI retries: 2 mitigation; hover-first fix 65a1ad7.
• Verify and fix fixed-node right-click release in native GPU Chromium #22 — right-click variant of the same deferred-hover mechanism (tests/e2e/pointer.ts).
• CI: Validation workflow takes ~15 min; shard the Playwright e2e suite to cut wall clock to ~5-6 min #30 — "do not trim qualified waits" constraint.
• Make the native-GPU e2e lane default to parallel workers (SwiftShader gate stays serial) #56 — GPU-lane parallel default (separate; unblocked by Firefox e2e flake: "background drag" rotation (matrix.spec.ts:224) — synthetic-input-delivery nondeterminism (survives on hardware) #55).
• tests/e2e/matrix.spec.ts (click-to-focus test), tests/e2e/pointer.ts, tests/e2e/graph-handle.ts, playwright.config.ts:116-132.
• Lesson (COLD, codev/resources/lessons-learned.md, added by Firefox e2e flake: "background drag" rotation (matrix.spec.ts:224) — synthetic-input-delivery nondeterminism (survives on hardware) #55): a hard-coded "background"/empty screen coordinate in an interaction test is not guaranteed to stay empty when the scene changes.