PermissionMatrixEditor artifact tier over-locks on single-kernel hosts (environmentId undefined) — accepted residue of #4518

[yinlianghui]

Observation-class residue recorded from #4518 / PR #4525, for triage. Not a defect in the landed card — a documented consequence of its own ruling.

What was landed: PR #4525 gives PermissionMatrixEditor the server's second write tier: at the environment door (no packageId), a code-declared (artifact-backed) permission set renders read-only unless allowOrgOverride, mirroring the server's NOT_OVERRIDABLE gate at protocol.ts:8888 and ResourceEditPage:1332's predicate.

The residue: the server's artifact gate only engages when the kernel has an environmentId (if (this.environmentId !== undefined)). On a single-kernel host that property is undefined, the server disengages the artifact tier entirely, and an env-scope edit of a code-declared set is accepted with 200 — while the editor now renders it read-only. Direction is conservative (an honest lock instead of a Save that would work), and nothing #4519 opened that was previously reachable regresses. But it is a real client/server divergence on that host class.

Why it could not be closed in #4518: the client cannot see the kernel's environment topology. useMetadataClient() takes no argument at all 30 app-shell call sites; MetadataClient bakes the scope into a private base with no getter; the only surface reporting the resolved environmentId is GET /discovery — a new server probe, which the card's ruling forbade.

Paths if triage wants it closed exactly (design decision, pick one):

1. Plumb the resolved environmentId through MetadataClient / discovery into the editor, and disengage the client tier when it is undefined.
2. Accept the conservative lock as permanent documented behavior for single-kernel hosts.

Refs: #4518, PR #4525, #4519, #4446. Filed unassigned for triage.

---

Generated by Claude Code

[hotlong]

Triage: finding — accepted, documented residue of #4518's own ruling, not a live defect (the client fails conservatively toward an honest lock, nothing regresses). Real divergence exists only on single-kernel hosts, where the server accepts what the client now blocks. Held for a future findings-triage round to grade; closure requires a design call the filer already scoped (plumb environmentId through MetadataClient/discovery vs. accept the lock as permanent documented behavior for single-kernel hosts) — that call should go to needs-user-decision if and when this gets promoted, since #4518's ruling already forbade adding a new server probe for this once.

---

Generated by Claude Code