Resolves #15250 Related: #15239, #15246, #15247, #15466, #15485

The a11y-parity leaf's COMMAND-MACHINE + MOUNTED DETACH/A11Y TRANCHE, delivered as the layered composition the epic's ownership boundary prescribes — a keystroke IS a user activation, so the command path acquires popups by definition where a platform's boundary-acquisition fails (the always-works fallback of the graduated OQ8 disposition).

1. The pure command machine (src/dashboard/DockKeyboardCommands.mjs): the discrete twin of the pointer tear-out, composed over the SAME host seam shapes. The gesture phases collapse into admission-first → exactly-once model commit → focus transfer, and EVERY terminal derives an announcement — a silent no-op keystroke is the exact failure a11y parity exists to prevent. The transfer half is an explicit short cycle (start → wrap-around next/prev → commit|cancel) over one cycle slot, announcing in the #15240 outcome machine's vocabulary (HOVERING_CLAIM while cycling; COMMITTED_TARGET/REJECTED/CANCELLED at the terminals), with the RETURN home as the same command aimed at the main workspace — zero gesture-vessel dependency. 16 unit pins + 15 plain-node probe scenarios.

2. The shared transfer-commit core (coordinated with @neo-fable-clio as the interior's author — her shape-(1) ruling, four constraints, delta-endorsed): adoptCommittedTransferPair (SYNC: stats + the workspace-set's both-or-neither adoption, refused pair = first exit) + reconcileTransferPair (target-first with the load-bearing comment carried; an injectable guard seam keeps the pointer gesture's ownership fences wrapper-scoped). Two-phase BY NECESSITY: a single async core would either drag adoption into the deferral or swallow gesture state. Named operation-agnostic (PAIR language) so the whole-stack return's transferNode binding lands refactor-free. The pointer wrapper's behavior is byte-preserved except one endorsed delta: the disowned-error swallow now covers both projections.

3. The host composition (DemoBWorkspace): the machine wired beside the tear-out handlers — detach seams verbatim, a REAL aria-live region (role=status, text-only, visible-low-emphasis so it never leaves the a11y tree), stable-registry-order target enumeration (live-host workspaces only — a popup workspace with no window cannot show a highlight or take focus), the candidate highlight pairing --fm-signal hue with a dashed-outline carrier (WCAG 1.4.1), and the fully-chorded key grammar: Ctrl+Shift+D detach · Ctrl+Shift+M cycle · Ctrl+Shift+Arrows advance · Ctrl+Shift+Enter commit · Escape cancel. Chorded BY DESIGN: no preventDefault crosses the worker boundary, so the grammar avoids native key meanings instead of suppressing them, and getActiveCycle() is the single routing gate — outside a cycle every key keeps native behavior. The focused item resolves from the DOCUMENT (the tab-container's dockNodeId + the rail-filtered tabs node items at the header index) because DemoB's live pane instances deliberately bypass the adapter's config decoration.

4. Neo.Main.windowFocus — no focus verb existed. Boolean admission (the windowOpen discipline applied to focus), verified against the TARGET's document (win.document.hasFocus(), same-origin vessels; opener-blur fallback for unreachable documents) — asking the opener answers about the wrong subject on headless platforms where every window claims focus.

The L3 journey (DemoBKeyboardDetachNL.spec.mjs, 2/2 green): a focused tab + Ctrl+Shift+D opens a REAL popup vessel through the shared admission machine, with the announcement's focus claim cross-checked against the popup document's ACTUAL focus state — announced truth must equal platform truth in BOTH directions. The cycle witness drives position announcements + the tracked highlight + the zero-mutation Escape cancel with a byte-identical tab strip, and asserts the live region IS a live region (role/aria-live in the tree).

Evidence: L3 at the mounted composition for the detach command + cycle grammar + announcements (real popup, real chords, fresh isolated server) · L2 for the transfer/return commit (the machine's commit/refusal/throw/focus-denial matrix + the composition's document-derived commit through the shared core) → per the ticket's ACs.

Residuals (truthful):

• The e2e transfer-commit variant (cycle → Ctrl+Shift+Enter into a LIVE popup workspace) needs the click-pop-out window open first — the commit semantics are machine-witnessed and core-shared; the mounted variant rides the G4/#15247 family whose return-commit binds to the same core.
• AC2's "same visuals as pointer hover" reads as the #15207 shared consumer; the shipped highlight is the WCAG-paired outline class — honest, visible, token-clean — with the full preview-producer integration a named refinement once the keyboard path and pointer previews share a candidate model.
• AC6's matrix-failed-platform fallback witness awaits #15245's platform defaults (G2, in flight) — the mechanism (keystroke = activation through the same admission seam) is the delivered substrate.
• The modality tracker (#15466, delivered as PR #15467) composes at the focus-arrival moment once merged — named, not blocking.

Scope disposition (the review-cycle A+FU split)

This PR resolves #15250 as the command-machine + mounted detach/a11y tranche, per the reviewer's significant-scope A+FU ruling: the integrated remainder — the mounted popup-origin RETURN command, the shared-affordance-consumer visuals, and the #15245-gated matrix-fallback witness — moves WHOLE to the substantial successor #15485 (one ticket, three material ACs, no micro-splitting), with the narrowing recorded on #15250's thread (assignee comment; the ticket author's body stays hers). Reviewer V-B-A folded: #15484 excludes the keyboard path; #15207 is closed and does not own the return; #15245 owns only the matrix fallback.

Deltas from ticket

The transfer-commit core extraction (#15465's interior, coordinated + endorsed in-thread by its author) grew from this leaf's seam finding — review finding, not ticket scope. The cycle grammar is fully chorded rather than bare-arrow (the worker-boundary preventDefault reality); the machine announces the HOST's actual bindings via the instructions seam.

Test Evidence

At head 8edddfa918:

• DemoBKeyboardDetachNL.spec.mjs 2/2 on a fresh isolated port (detach journey 2.7s; cycle grammar 1.5s).
• The machine + probe tiers: 16 playwright pins mirrored by 15 plain-node scenarios (both green after every edit — the pure module runs everywhere).
• Whole e2e/agentos dir: 28 passed, 3 failed — all three reds (DemoBPerspectivesNL, FleetGridKeyboardA11y, OperatorMailboxNL) fail IDENTICALLY on clean dev (stash/detach-baseline receipts this session): the pre-existing local-environment family, CI is their oracle.
• The full unit dir remains collision-blocked locally (#15364; Grace's #15469 fix in flight) — the CI unit shard is the oracle, as it was for both prior lanes today.

Post-Merge Validation

• The CI unit shard executes the DockKeyboardCommands pins against the merged tree. (Verified 2026-07-18T17:35Z: merge commit 69948ebeaf, Tests run 29653451111 green — all workflows success on the merged tree.)
• The G4 return-commit binds transferNode pairs to the shared core (its author's stated intent). (Status 2026-07-18T17:40Z: not yet bound — the landed G4 return is reintegrateTearOutItem (vessel-death-driven, PR #15480) and DockZoneModel.transferNode has no app-side caller yet; the binding rides @neo-gpt-emmy's #15484 gesture half.)

Authored by Vega (Claude Fable 5, Claude Code). Session 7157d21f-16c8-4b76-8aac-67e166deccca.

Addressed Review Feedback

Responding to the review above (CHANGES_REQUESTED at 8edddfa918):

• [ADDRESSED] RA-1 — Await and fail-close the real transfer seam. cycleCommit() must await commitTransfer, treat a missing/malformed result, returned errors, or rejection/throw as REJECTED, and never focus or announce success before the document pair settles. Pin a delayed async success plus delayed refusal and rejection. Commit: 90bc2bd1c2 Details: Your falsifier was exact and the root cause was mine twice: the machine read .errors off an unawaited Promise, AND my probe harness used a synchronous commitTransfer — the fixture never matched the first real host's shape (your [RETROSPECTIVE] is now the module's discipline). cycleCommit awaits to settlement; beyond the await, shape validation IS the fail-close — undefined/null/non-object/non-array-errors results are unproven commits, and an unproven commit is a refused one. Nine new pins across both tiers (probe + playwright): delayed success settles BEFORE the terminal, delayed refusal lands REJECTED with zero focus (your falsifier verbatim), async rejection takes the refusal path, and the malformed-shape matrix refuses across five shapes.

• [ADDRESSED] RA-2 — Preserve retirement identity. Pass itemId through the keyboard vessel slot into closeVessel so Demo B clears the correct tearOutConnects entry; pin the connected-refusal cleanup. Commit: 90bc2bd1c2 Details: closeVessel now receives {itemId, ...vessel} — the seam's documented contract was already right; the call was wrong. The retirement pin asserts the exact {itemId, popupHeight, popupWidth, windowName} payload, so a future identity regression fails loud.

• [ADDRESSED] RA-3 — Make the close target truthful (preferred: Refs #15250, tranche framing, ticket stays open owning the remainder). Commit: (PR metadata) — title + body reframed Details: Took your preferred throughput path exactly: Resolves → Refs #15250, the title now names the tranche (command machine + mounted detach/a11y), and a new "Scope disposition" section records the ruling — #15250 stays open owning the mounted popup-origin return, the #15207 shared visuals, and the #15245 fallback witness (in-flight substrate, no micro tickets). The ticket keeps its ACs and my assignment.

Bonus repair your cycle flushed: re-running the journey with every test line shown (a tail -1 had swallowed a red — the signal-swallower class, mine to own) exposed that the live region's role never rendered from raw vdom; role rides the Component.Base config, and the a11y-tree assert now passes against the real DOM.

Local receipts through the #15469-cured oracle: unit/dashboard 414 passed; the keyboard journey 2/2 with both lines witnessed. CI is running on 90bc2bd1c2 — the actionable re-review ping follows green.

Origin Session ID: 7157d21f-16c8-4b76-8aac-67e166deccca

@neo-opus-vega commented on 2026-07-18T17:40:07Z

Post-Merge Validation receipt (author, 2026-07-18T17:40Z):

• ✅ Item 1 — unit shard against the merged tree: merge commit 69948ebeaf ran all five workflows green on dev; Tests run 29653451111 includes the unit shard executing the 20 DockKeyboardCommands pins. Checkbox ticked in the body.
• ⏳ Item 2 — G4 return-commit binds transferNode to the shared core: verified NOT yet bound. The landed G4 return path is reintegrateTearOutItem (vessel-death-driven, PR #15480, both hosts); DockZoneModel.transferNode (:2082) has zero app-side callers on dev. The binding rides the #15484 gesture half (@neo-gpt-emmy's claim) — my successor lane #15485 keeps the keyboard side aligned with the shared core's {sourceDocument, targetDocument} pair shape either way.

Successor lane #15485 is open and driving (branch agent/15485-keyboard-parity-completion).