A grid without a size constraint wedges the App worker and pushes the layout

Context

Found in manual testing of apps/workstation on dev (2026-10-22). A few steps after a fresh load (the history shows 3), the 100k Matrix pane runs past the right edge of the window and pushes its neighbours off-screen. Its "Living signal" sparklines stop rendering.

The App worker console repeats without end:

E [Base.mjs:32:17] Trying to create an item with an already existing id
(2) [Row, Row]

Severity: left running, Chrome climbed to about 160 % CPU and then the tab crashed ("Aw, Snap!"). The loop takes the whole app down, not just the pane.

@tobiu's screenshot follows in a comment. The three steps before it are not known yet.

Measurements

Live, the operator's session, read-only over the Neural Link:

• The bridge's 1000-entry log buffer for this worker is a single 44 ms burst, about 25 s after load: 500 × the duplicate-id error, each paired with its [Row, Row], for ids neo-grid-body-1__row-559 … __row-808.
• From then on, every Neural Link call to that worker timed out (five different calls), while the bridge itself stayed healthy.
• Row ids are ${bodyId}__row-${rowIndex % poolSize} with poolSize = availableRows + 2 · bufferRowRange (grid.Body#getRowId). An id of 808 therefore means a pool of at least 809 rows. That puts availableRows near 800, where the same pane normally holds 30. The body had measured an unconstrained height.

Headless, my clone (8bb3574d0f, whose src/ matches dev): in the shipped layout, the Matrix pane's chain is made content-sized: height: auto; flex: none on its tab container, tab body, data pane and grid container.

That is 2 of 2 runs. The forced variant reproduces the hang and the layout push, but not the duplicate-id loop the live session showed.

The Problem

Two engine defects, plus one unknown trigger:

1. An unconstrained container wedges the grid. Once its container is content-sized, the grid's own output feeds its next measurement. The App worker stops responding and the pane grows past its dock slot. Which function spins is not yet traced. The resize-echo classification from #16375 does not cover this path.
2. Pool growth collides with live ids. Body#createRowPool() sets rowPoolSize to the new size, then Neo.creates rows with getRowId(current + i). When the pool shrinks and regrows, or when recycled items no longer sit at the slot their id names, a new row can carry an id that is still registered. manager.Base logs it, and the next pass tries again, which gives the endless [Row, Row] loop. The live session shows the loop; which of the two paths produces it is inferred, not traced.
3. Trigger unknown. Something in the operator's three steps left the pane without its size constraint. Once it is known it gets a reproduction, and if it is a dock defect, its own ticket.
   • Suspect checked and ruled out on its own (#19992): after the crash, the restored layout from that session has its edge-zone root wrapped in a vertical split by a root-edge drop. Headless, committing that exact document through the Group (onDockZoneDocumentChange(document, {operation: 'applyDocument', document})) and then reloading kept the worker healthy: it answered in 3–4 ms, logged no duplicate ids, and held 31 Matrix rows. The wrapped root alone is not the trigger, which stays open.

Measured at intake (comment 6339978228), which supersedes the inferred parts of 1 and 2:

• Under the forcing, the view measures the body's full scroll height (5,000,050 px), and createRowPool creates 99,970 rows.
• items.push(...newRows) then throws a RangeError (Body.mjs:736) after every row has registered its id. Each later pass re-creates those ids, which produces the [Row, Row] flood.
• Neither inferred collision path (shrink then regrow, regrow while scrolled) reproduces.

The Architectural Reality

• src/grid/Body.mjs: createViewData() maps the mounted rows onto the pool, createRowPool() grows by Neo.create and shrinks by pop() + destroy(), and getRowId() is the ordinal pool id.
• src/manager/Base.mjs:32: the duplicate-id refusal that logs the pair.
• The grid's measurement path (grid/Container, layout-box reads, and the resize-echo classification from #16375).

The Fix

• Bound the work by the screen. Clamp availableRows, and with it the pool, to what the window can actually show, so an unconstrained container cannot ask for hundreds of rows.
• Collision-free pool changes. Never Neo.create a row whose id is registered. Reuse it, or rebuild the pool atomically when its size changes.
• Break the feedback. A size change the grid caused through its own content must not re-trigger its layout. Extend the echo classification, or measure against the dock slot rather than the content.
• The trigger, once the operator's steps are known: fix it where it lives.

Acceptance Criteria

• E2E: a new whitebox Neural Link spec, test/playwright/e2e/workstation/WorkstationGridUnconstrainedNL.spec.mjs, makes the 100k Matrix pane's chain content-sized as in the probe. Within 1 s it asserts four things: the App worker answers in under 100 ms, the view and the row pool fit the window, rows still render, and no duplicate-id error is logged. It must be red on dev (worker unresponsive) and green with the fix, and it joins the hosted Workstation NL tier (#19655).
• Unit: growing the row pool past the call-argument limit leaves every created row in the pool, never registered outside it. It must be red on dev (the RangeError). Shrink then regrow, and regrow while scrolled, keep every row id registered once, as regression guards.

Out of Scope

• #19985, the same grid's drag-proxy skin.
• Content-sized rows (#19658), an opt-in mode this pane does not use.
• The trigger (the operator's three steps), and the pane widening under it. Any content-sized chain takes its content's width, so the widening is the trigger's symptom, not the grid's. #20009 owns the trigger. Split at intake, so that this ticket closes on the two engine defects.

Related

• #16375: transform-immune measurement and resize-echo classification.
• #19655: the hosted Workstation NL tier.
• #19982–#19990: the same testing pass.

Live latest-open sweep: checked latest 20 open issues at 2026-10-22T10:37:14Z; no equivalent found. Keyword sweep: "already existing id grid row", "grid row pool duplicate", "grid resize loop worker", "grid content-sized container"; closed adjacent only (#19658, #19668, #18762, #18755). A2A claim sweep: no overlapping lane-claim. MC sweep: "grid row pool trying to create an item with an already existing id…", 6 results, no prior decision (a 2026-08 note confirms pool-ordinal ids). Own-assignment sweep: none overlapping.

Origin Session ID: 70e59a84-efd6-4e7a-9262-c5478bd314ff Retrieval Hint: "grid unconstrained container App worker hang, row pool duplicate id error loop, Workstation 100k Matrix layout push"