LearnNewsExamplesServices
Frontmatter
id18696
titleThe dense tour counts every root `state.Provider`, so the topology Group''s history Provider reads as a second one
stateClosed
labels
bugaitestingdashboard
assigneesneo-opus-grace
createdAtSep 14, 2026, 3:37 AM
updatedAtSep 14, 2026, 7:58 PM
githubUrlhttps://github.com/neomjs/neo/issues/18696
authorneo-opus-grace
commentsCount0
parentIssuenull
subIssues[]
subIssuesCompleted0
subIssuesTotal0
contentTrust
projected
quarantined0
signals[]
blockedBy[]
blocking[]
closedAtSep 14, 2026, 7:58 PM

The dense tour counts every root state.Provider, so the topology Group's history Provider reads as a second one

Closed v13.2.0/archive-v13-2-0-chunk-21 bugaitestingdashboard
neo-opus-grace
neo-opus-grace commented on Sep 14, 2026, 3:37 AM

CORRECTED 2026-09-14 after a captured construction stack refuted this ticket's own premise. The title and the original body claimed a Provider "bound to nothing". It is bound to a Group. The correction is kept in place rather than rewritten away, because the wrong reasoning is the useful part.

Context

WorkstationNL › the real tour keeps density, data, Canvas output, themes, rails, and identities live fails at its very first identity assertion:

Error: Workstation owns one root StateProvider
Expected: 1
Received: 2

This red rode #17844's census for several runs recorded as "flagged to #18461" and read as owned. #18461 has since closed without touching it, so that pointer was dead and the red had no open owner. This ticket is that owner. The census body has been corrected.

What it actually is

Neo.manager.Transaction#getProvider (src/manager/Transaction.mjs:537) creates one state.Provider per Group, on first read — deliberately parent-less and component-less — so a popped-out window and its opener read the same canUndo / canRedo answer. retireGroup destroys it with the Group (:915). It is a legitimate second root.

The Workstation reaches it at boot because the topology bar's undo/redo carry history badges (#18560); TopologyToolbar.historyStepBadge calls getProvider inside a binding formula, and #18677 only moved that binding into its own view.

How that was established, since the first answer here was wrong. A temporary warn in Provider#construct, firing for any instance built with neither component nor parent, named the creator:

Provider.construct                src/state/Provider.mjs
Object.create                     src/Neo.mjs
Transaction.getProvider           src/manager/Transaction.mjs:537
TopologyToolbar.historyStepBadge  apps/workstation/view/TopologyToolbar.mjs:98

and TransactionManager.getProvider(workspace.topologyGroupId).id then resolved to that exact instance, neo-state-provider-2.

The refuted hypothesis, kept

The original body argued the orphan was very likely a destroyed Provider still registered: a probe had returned its id and apparently nothing else, which is core.Base#destroy's exact fingerprint (_id survives the delete loop). Two things were wrong with it.

  • The reasoning was a reconstruction from reading destroy, afterSetId and beforeSetStateProvider — and a reconstruction cannot refute the hypothesis it was built from. The stack could, and did.
  • The discriminator did not discriminate. isDestroyed reads undefined on the two healthy Providers too, so its absence on the orphan was never evidence of anything. Asserting from it would have been a control on the wrong axis.

There is no leak and no engine defect here. Neo.manager.Transaction is correct as written.

The Fix

The arm is what is wrong: it counts roots globally and assumes only the view tree makes them.

  1. Count the root Providers that own a component — one, the workspace's, which is what the identity snapshot means — instead of every parent-less instance.
  2. Name the component-less root by id, from the manager, rather than relaxing the count to 2.

Two reds were hiding behind it

The count aborted the arm at :664, so nothing downstream of it had a run for as long as #18560 has been in. Both are newly visible, not newly broken:

  • :876 shortGeometry.scale.height >= 230 — in scope here, because it is the same defect shape (an assertion stating what the composition does not owe) and it sits between :664 and any receipt that the repair holds. At a 600-tall viewport every band keeps its rem floor — asserted two lines above — so the center absorbs the whole block-size delta: the grid goes 516 → 216 for a 300-pixel shorter host. Paying 230 out of 216 requires pushing the bottom band under that floor. The two assertions cannot both hold.
  • :1266 the dark-mode splitter's element.matches(':active') after mouse.down() — out of scope, noted in #17844. Independent of this ticket's subject, unowned, and mode-independent (it fails headed and headless alike, so it is not the #18687 shape). Possibly the same ground as #17889.

Acceptance Criteria

  • AC-1 The dense tour's root-Provider assertion passes against a booted Workstation, by counting component-owning roots — not by relaxing the expected count.
  • AC-2 A component-less root is asserted by identity: the arm resolves the topology Group's history Provider from Neo.manager.Transaction and requires it to be exactly the set of component-less roots. A second, unexplained one still fails the arm.
  • AC-3 shortGeometry.scale.height is asserted as the viewport-delta relation it actually is, with a floor that can hold at 600 CSS pixels.
  • AC-4 Both repaired assertions behave identically headed and headless.
  • AC-5 #17844's census records this arm's disposition and carries :1266 forward as an attributed, unowned red.

Out of Scope

  • :1266, and every other red in #17844.
  • Any change to Neo.manager.Transaction, which this ticket set out to fix and should not touch.
  • Whether the arm should count roots at all. It should; the count is what surfaced all of this.

Related

  • #17844 — the census carrying this red, corrected to show it unowned rather than flagged to a closed ticket.
  • #18461 — closed; was cited as this red's owner and never touched it.
  • #18560 — gave the topology bar its history badges, and with them the second root at boot.
  • #18677 — moved that binding into TopologyToolbar; not the cause.
  • #17889 — pointer-capture migration, a candidate ground for :1266.