Resolves #14658

The FM cockpit's live half of the §01 mission-control layout: the committed dockZone.v1 document (the landed cockpitDockDocument data leaf) becomes the cockpit's layout SSOT, projected through DockLayoutAdapter with the proven reducer / view-sync commit loop — splitter drags, cross-zone tab drops and NL-driven operations all funnel through ONE pure reducer (DockZoneModel.applyOperation) and ONE deferred re-projection (me.timeout(0) + the isDestroyed guard — the pattern source's documented use-after-destroy trap, kept load-bearing).

Design decisions (recorded per the epic's delta discipline):

• FleetCockpit itself becomes the projection host — no new orphan component. The cockpit already owns the provider (the roster store's sharing scope), the controller, and the live-feed loaders; the projection replaces its direct vbox arrangement, and the SSOT §01 split (~1.55fr fleet over 1fr activity, activity docked BOTTOM) is now EXPRESSED by the document's primary-split sizes rather than hand-set flex values. The record-driven substrate (grid/cards/stream) is wrapped, not re-laid-out.
• Runtime pane state lives on the OWNER, never on instances. Re-projections rebuild pane instances (removeAll + add), so resolveDockComponentRef re-materializes each pane from held state (gridAdapterState / streamAdapterState / streamEvents) — a committed layout change can never reset a live grid back to its sample badge. The loaders write both the live instance and the held state.
• Panes stay layout-blind (the docking pane contract): the resolver threads ordinary configs plus the stable FLIP marker class only — spec-asserted that no dock wiring reaches a pane config.
• agent-detail / perspectives render honest labelled placeholders behind their auto-hide rail tabs — their real views are sibling leaves (#14608 / #14616); a blank pane masquerading as a finished surface would violate the cockpit's honesty bar.
• The dock-holder contract ships complete: getDockZoneDocument() (the NL read half) + applyDockZoneOperation (the write half) + the cross-zone drop reducer (producer → previewToOperation → the same loop) — a human gesture and an NL operation are the same commit path, which the e2e proves literally.

Two integration defects found and fixed by building the loop:

1. The agentos app never declared the DockFlip main-thread addon (neo-config.json). A worker-side remote call into an addon the app never loaded does not reject — it never settles — so the view-sync's await captureFirst(...) silently hung the ENTIRE re-projection loop with no error surfacing anywhere. Fixed (addon declared), and the trap is now documented in the adapter's consuming-workspace JSDoc beside its existing theme-file requirement (src/dashboard/DockLayoutAdapter.mjs — the only src/ touch, doc-only).
2. The wholesale-refresh reconciliation defect (#14985) reproduces deterministically on the cockpit: after a committed drag, the worker provably rebuilds (breadcrumb-instrumented: scheduled/fired/entered/completed all 1, no errors; new splitter instances exist) while the DOM keeps the stale pre-commit tree. This is the already-ticketed framework recurrence (its owner's broadcast lists dock surfaces as casualties) — NOT introduced here. The e2e therefore asserts post-commit truth WORKER-SIDE (instance identity across re-projection via the NL — the truth layer the Neural Link exists for); the DOM-flush regression witness belongs to that fix.

Evidence: L3 (the live commit loop driven on the mounted cockpit: real pointer drag at real cadence + NL operation, both committing and re-projecting — plus the full agentos e2e suite) → L3 required (the AC names NL-verified drag on the mounted cockpit). Residual: the DOM-flush witness rides #14985 (framework defect, owned out; worker-side truth fully proven).

Deltas from ticket

• The DockFlip addon declaration + the adapter JSDoc trap note (integration defect 1 — found by the falsifier, required for the loop's view-sync to run at all).
• The cross-zone drop reducer rides along (the ticket names the resize loop; the projected tab affordances exist either way, and a silently dead drop path on a mission-control surface would be an incomplete contract — same loop, one more gesture).
• The e2e's post-commit witnesses are worker-side (see defect 2) — the ticket's "NL-verified" phrasing is satisfied more literally than planned.

Contract Ledger

Test Evidence

At head 159996d3f:

npm run test-unit -- test/playwright/unit/apps/agentos/ --workers=1
193 passed / 1 failed — the red is fleetGrid.spec.mjs:103, PRE-EXISTING on clean dev (verified: identical failure with zero changes present via git stash -u; passes in isolation AND paired with the new spec file — a sweep-order heap-bleed from an earlier file, flagged to its owner)

NEO_E2E_PORT=8117 npx playwright test agentos -c test/playwright/playwright.config.e2e.mjs --workers=1
5 passed — the FULL agentos e2e suite including the new FleetCockpitDockNL journey (initial projection renders, read half serves, real drag commits + worker re-projects, NL operation round-trips) and all sibling surfaces (Cockpit, Accounts, both lifecycle NL specs) against the converted cockpit

The six new unit specs: reducer purity + fail-closed rejection, view-sync stores-synchronously/re-projects-deferred, the destroy guard, behavioral instance-binding of threaded callbacks, layout-blind panes + owner-held state re-materialization, both live panes projected exactly once (auto-hidden chrome stays rail material).

Post-Merge Validation

• When #14985 (the wholesale-refresh reconciliation defect) closes, add the DOM-flush geometry witness to FleetCockpitDockNL.spec.mjs (splitter position after commit) — the worker-side witness stays either way.
• The fleetGrid sweep-order red (pre-existing, isolated-green) needs its owner's disposition — repro triple documented above.
• #14616 (perspectives pane) consumes the landed DockPerspectiveStore + replaces the placeholder; #14608 (agent detail) replaces its placeholder.

Process note: authored during the operator-granted temporary Fable 5 window.

Authored by Grace (Claude Fable 5, Claude Code). Session ef6b9a4a-54ec-4afb-8438-f89a3ee46ad2