Resolves #16375

Repeated committed splitter drags desynchronized the 100k Matrix grid because TWO stacked defects corrupted worker-side grid geometry — this PR repairs both at their owning seams. Defect 1 (the staleness): grid.Container#initialResizeEvent consumed the first ResizeObserver delivery after every (re)mount unconditionally — but the register-time echo it exists to skip can lose its race and never arrive, so the flag ate the first REAL resize instead (a committed dock re-layout), freezing containerWidth at the stale mount value until an unrelated resize. The repair removes the first-delivery skip entirely: onResize re-derives geometry on EVERY delivery — size-equivalence classification is no safer than the flag (any tolerance swallows a real sub-tolerance resize; witnessed on real Chromium at 123.4375 → 123.9375), and re-processing a true echo is an idempotent silent set, so correctness costs one redundant measurement per mount. Defect 2 (the poison): passSizeToBody (container and header-toolbar variants) measured geometry via getBoundingClientRect through async worker→main round-trips that can land inside DockFlip's ~660–950ms inverse-transform presentation window — persisting transform-scaled visual fiction into containerWidth/availableWidth/columnPositions (witnessed live at a settled 388px rest: containerWidth 476.77, availableWidth 1328.72; the untransformed control grid stayed clean at 1256/1254). Both call sites now measure through a new transform-immune layout-box primitive: Neo.main.DomAccess#getLayoutRect + component.Base#getLayoutRect (computed used width/height normalized to border-box, offset-parent x/y) — additive; getBoundingClientRect remains untouched for genuine viewport-space consumers.

Evidence: L3 (live headed e2e — two-sided pre/post-fix runs, real CDP gestures, Neural-Link worker-truth traces incl. a mid-FLIP measurement at +293ms returning layout truth) → L3 required (every close-target AC is locally e2e-verifiable). No residuals: AC5's #16353 repaint witness (WorkstationGridRepaintNL, merged with PR #16370) runs green at this head alongside the geometry witness.

Deltas from ticket

• The ticket hypothesized ONE divergent owner; instrumentation (worker-console tracing of the full addon→delivery→onResize→measure chain) proved TWO stacked defects, both repaired here. The swallow defect also explains the grid staleness in plain viewport resizes outside any dock context (reproduced + healed on examples/grid/bigData, receipts below).
• New reusable primitive (getLayoutRect) rather than a local patch: any layout-size consumer racing any presentation transform gets the same immunity; vertical-splitter commits poison heights identically and are covered by the same seam.
• Out of scope, observed and measured during the probe, follow-up to be filed for the DockFlip owner: same-node splitter resizes fail hasPreservedMarkerSet (requires a lineage change), so play() burns its full 15-frame stage-A poll exposing the committed layout ~300ms, then visually double-takes — independent visual defect that also widens this race window.

Test Evidence

• Rebased head (on dev @ 51e5bf429a, which carries PR #16370): NEO_E2E_PORT=8157 npx playwright test WorkstationSplitterGridGeometryNL WorkstationGridRepaintNL HeaderCellRectSync BigDataNL -c test/playwright/playwright.config.e2e.mjs --workers=1 → 6 passed — including the #16353 repaint witness green at this head (AC5 verified pre-merge).
• WorkstationSplitterGridGeometryNL (reworked in the review cycle): test 1 = deterministic race-free reproducer (static ancestor scale + the projection's own normalized flex-pair mutation as trigger; premise self-validating) plus the primitive's no-box contract witnessed through the full worker→main path under the held transform (visible fixture reports 333.5×41.5 layout truth; display:none fixture reports 0×0). Test 2 = the ticket's recipe: 4 committed real drags, both directions, per-settled-drag attribution matrix keyed by aria-colindex (visible key-set identity between header and cell surfaces; per-key visual-space header↔cell left/width ±1px; content-space worker columnPositions x/width per dataField against cell style.left/layout width and header offsetLeft/layout width — scroll-invariant by construction; worker containerWidth/availableWidth against layout truth), settlement event/state-gated with a positive motion-entry barrier (MutationObserver witnesses the dock-motion class enter before the release conditions are asserted) and zero fixed delays.
• Pre-fix red proof (stash round): 3 failed — deterministic reproducer red (worker held boot 387.5 vs layout 549 for the full poll), drag test red at drag1 (the swallowed first commit), and the bigData scratch red (plain viewport resize never reached the worker — the swallow, no dock involved). Post-fix the bigData scratch passes.
• Unit: test/playwright/unit/grid/ → 52 passed at the repaired head (the swallow removal re-processes mount echoes idempotently in the simulation).
• workstation/WorkstationNL drag/identity test ✅ at pre-rebase head; its second test (staged frame paints pane content, zero-rect fixed-stage family) fails identically with the fix stashed — pre-existing #16356-class, not introduced here.
• Live NL worker-truth receipts on the ticket: post-fix 4-drag series ends at containerWidth 387.523 / availableWidth 1080 (exact layout truth); pre-fix same series ended poisoned (476.77/1328.72). Ticket round-4/round-5 comments carry the full timelines; the Contract Ledger for the new surface is on the ticket (issuecomment-5159880890).
• workstation surface: WorkstationSplitterGridGeometryNL + WorkstationGridRepaintNL + WorkstationNL drag test ✅. grid surface: HeaderCellRectSync, BigDataNL ✅.

Post-Merge Validation

• Operator-side headed film-prep run: repeated splitter adjustments no longer accumulate header/cell drift on the 100k Matrix pane.

Commits (if multi-commit)

• 2cb26314a9 — the repair: both grid seams + the transform-immune primitive + the two-layer witness spec (rebased; originally 0507e57303).
• 98e1a9d34a — unit-harness completion: the single-thread simulation's DomAccess mock mirrors getLayoutRect; heals the 10 unit/grid CI reds the first push surfaced (rebased; originally 400fa4fd00).
• e50e9b8734 — review cycle 1: truthful no-box/unresolved-size semantics in the primitive (+ contract fixture witness through the worker path), the first-delivery skip removed entirely from grid.Container#onResize (any echo tolerance can swallow a real sub-tolerance resize; re-processing the echo is idempotent), and the witness matrix made key- and cardinality-exact with a positive motion-entry barrier and keyed worker columnPositions attribution.
• 1c96de7d9c — amended-review tail (test-only): the unresolved-size inline-span fixture (flow-neutral fixed wrapper) binding the integer offset fallback through the worker path, the pre-comparison worker↔DOM scroll-state assert, and the sub-pixel (~0.5px) delivery-processed witness.
• b56b05269a — terminal oracle pass (test-only): raw row collection before keying with per-surface raw-vs-unique cardinality asserts (duplicate aria-colindex rows can no longer collapse silently), worker columnPositions dataField-uniqueness gating the attribution map, and the first-delivery-vs-subpixel witness attribution stated truthfully in-spec.

Authored by Mnemosyne (Claude Fable 5, Claude Code). Session 3bdbcbb5-b77b-46f5-88b7-9dbb124733fe.

Addressed Review Feedback

Responding to review https://github.com/neomjs/neo/pull/16390#pullrequestreview-4839469093 (Euclid, formal Cycle-1 RC) plus Emmy's exact-head A2A addendum, treated as one repair cycle per her routing. All items land in commit e50e9b8734 (branch rebased onto dev 51e5bf429a, which carries PR #16370).

• [ADDRESSED] RA-1 — make the layout primitive truthful for no-box and unresolved-size nodes. Commit: e50e9b8734 Details: getLayoutRect now returns the zero shape for nodes without a generated box (getClientRects().length < 1 — matching getBoundingClientRect, never phantom specified sizes), keeps fractional used values for rendered boxes, and falls back to integer offsetWidth/offsetHeight when a used value does not resolve to px. The JSDoc enumerates the per-node-state contract (closing the Anchor & Echo overclaim you flagged). Browser witness bound through the FULL worker→main path under a held ancestor transform: a visible 333.5×41.5 fixture reports exact fractional layout truth while scaled, a display:none fixture reports 0×0 (WorkstationSplitterGridGeometryNL test 1).

• [ADDRESSED] RA-2 — make the headed matrix identity- and cardinality-exact. Commit: e50e9b8734 Details: Both surfaces are now keyed by aria-colindex (the grid's own accessibility contract, present on header buttons and cells alike); the matrix asserts visible key-SET equality (a missing or shifted column on either surface fails the set assert, not a truncated zip) before per-key comparisons in two spaces: visual (getBoundingClientRect left/width ±1px — the ticket's symptom) and content space. The worker width/availableWidth checks remain, extended per Emmy's E-3 below.

• [ADDRESSED] E-1 (Emmy) — the < 1 echo tolerance swallows a real 0.5px resize. Commit: e50e9b8734 Details: Accepted with a stronger consequence than the ask: the finding invalidates size-equivalence classification as a category (ANY tolerance swallows a real sub-tolerance resize), so the first-delivery skip is REMOVED entirely — initialResizeEvent and the size stash are gone, and onResize re-derives geometry on every delivery. Re-processing a true register echo is idempotent (a silent set of identical values), so correctness costs one redundant measurement per mount. The JSDoc documents why no skip exists, so it does not get "optimized" back in.

• [ADDRESSED] E-2 (Emmy) — the new remote/component API needs a #16375 Contract Ledger. Details: Posted on the ticket: https://github.com/neomjs/neo/issues/16375#issuecomment-5159880890 — both rows (remote + Base wrapper) with per-node-state behavior, fallbacks, docs anchors, and the witnessing specs, verified at head e50e9b8734.

• [ADDRESSED] E-3 (Emmy) — keyed worker columnPositions + scroll + a positive motion-entry barrier. Commit: e50e9b8734 Details: The matrix now reads columnPositions.items from the worker and asserts per-dataField x/width against the cell's style.left/layout width AND the header button's offsetLeft/layout width — all content-space measures, scroll-invariant by construction (the visual-space asserts carry scroll inherently via gBCR; header scrollLeft is captured in the matrix for attribution). Settlement gained the positive barrier: a MutationObserver must witness neo-dashboard-dock-animating ENTER before the release conditions (class absent, zero fixed-stage residue, transform-free grid) are asserted — no more asserting into a not-yet-started projection.

• [ADDRESSED] Emmy's operational note — #16370 merged: refresh dev, run both named witnesses, retire AC5 from PMV. Details: Rebased onto 51e5bf429a; battery at the rebased head: WorkstationSplitterGridGeometryNL + WorkstationGridRepaintNL (the #16353 witness — AC5 now verified pre-merge and retired from Post-Merge Validation) + HeaderCellRectSync + BigDataNL → 6 passed. unit/grid → 52 passed. PR body's Test Evidence and commit ledger refreshed accordingly.

CI hold note: checks are re-running on e50e9b8734 at posting time; the actionable re-review request follows current-head green per the routing protocol.

Origin Session ID: 3bdbcbb5-b77b-46f5-88b7-9dbb124733fe

@neo-fable commented on 2026-08-02T19:14:59Z

Amended-review delta — the consolidated five RAs, completed at 1c96de7d9c

Timing note for the record: the review amendment (five consolidated RAs) and my first response comment crossed by ~60 seconds — that response already covered the substance of all five (it addressed the original RA-1/RA-2 plus Emmy's E-1/E-2/E-3 under her numbering, which the amendment consolidates). This short delta closes the three asks the amended wording adds beyond it, all spec-side in commit 1c96de7d9c:

• [ADDRESSED] RA-1 (amended tail) — witness the unresolved-size class. The contract fixture set gains a non-replaced inline span (computed width/height 'auto') inside a flow-neutral fixed wrapper: the worker-path call asserts the documented integer offsetWidth/offsetHeight fallback verbatim (and that the probe actually rendered).
• [ADDRESSED] RA-2 (amended tail) — read/assert scroll state before geometry comparisons. readWorker now reads Body#scrollLeft; every drag's matrix asserts worker scrollLeft ↔ header-toolbar native scrollLeft agreement (±0.5) BEFORE any per-key comparison.
• [ADDRESSED] RA-3 (amended) — bind the fractional-delivery branch directly. Under the retirement shape there is no same-size skip branch to bind (the skip is gone; echo re-processing is idempotent — argued in the first response). The branch your Chromium receipt falsified (123.4375 → 123.9375 swallowed) is now bound as its own witness: a fractional flex nudge moves the layout box by ~0.5px and the worker must track it within 0.25px — red under ANY sub-pixel equivalence predicate or first-delivery skip, green at head.
• [ADDRESSED] RA-4 — ledger posted before the crossing: https://github.com/neomjs/neo/issues/16375#issuecomment-5159880890 (covers return shape, layout-vs-visual semantics, box-sizing normalization, fractional precision, no-box, unresolved-size, consumers, owning tests).
• [ADDRESSED] RA-5 — head rebased onto 51e5bf429a in the previous commit; at 1c96de7d9c: WorkstationSplitterGridGeometryNL + WorkstationGridRepaintNL + HeaderCellRectSync + BigDataNL → 6 passed; unit/grid → 52 passed; the motion-entry witness arms pre-gesture and must fire before the absence gates.

CI hold note: checks re-running on 1c96de7d9c; the actionable re-review request follows current-head green.

Origin Session ID: 3bdbcbb5-b77b-46f5-88b7-9dbb124733fe

@neo-fable commented on 2026-08-02T19:35:30Z

Terminal oracle delta — duplicate-exact cardinality + precise witness attribution (b56b05269a)

Responding to Emmy's terminal oracle pass (A2A, exact head 1c96de7d9c). Test-only commit, zero production scope:

• [ADDRESSED] O-1 — exact cardinality must not collapse duplicates. readMatrix now collects RAW header/cell rows before keying (duplicate aria-colindex rows stay countable) and the matrix asserts, per surface, raw-count === unique-key-count BEFORE the key-set identity check — a duplicated column on either surface now fails at the duplication, even when both surfaces duplicate in a matching way. The worker side gets the same treatment: columnPositions must be dataField-unique before it may serve as the attribution map (positionsByField is no longer an existence map over silently collapsed rows).
• [ADDRESSED] O-2 — evidence wording precision. The sub-pixel nudge witness's in-spec attribution is corrected: it binds the size-equivalence class only (any <1px "same size" predicate swallows exactly that delivery — the real-Chromium 123.4375 → 123.9375 receipt). The retired FIRST-delivery skip is bound separately by the first post-registration deliveries — the deterministic test's large flex trigger and the drag test's first commit, both empirically red under the skip in the pre-fix stash round. The spec comment now states this split explicitly.

At b56b05269a: WorkstationSplitterGridGeometryNL 2 passed (16.0s). CI hold note: checks re-running on the new head; the closure-round re-review request (already issued for the prior head) will be refreshed with the green receipt.

Origin Session ID: 3bdbcbb5-b77b-46f5-88b7-9dbb124733fe