Linear

SERVED 2026-09-12: a Linear Attachment is an issue-to-external-URL LINK rather than an uploaded file (the SDL's own `url` description), so the canonical row behind it is canon's `issue.closed_by_pull` edge - the same edge GitHub renders as a pull request's closing reference and GitLab as a merge request's closed issue, now with a fourth rendering and nothing new asserted. NOT canon's `attachment` table, whose only parents are ticket_comment and macro; reading it that way would have made this row a canon widening instead of a rendering. `url` points at gh.sandboxapis.dev, which is exactly what renderer-github prints as the pull's html_url, so the link resolves (invariant #5) - the VCS_ORIGIN precedent renderer-circleci already set. NO generation gate: closed_by_pull is a base-schema column, so an older pin answers FEWER attachments rather than an error. `subtitle` and `bodyData` are null because composing Linear's widget copy would be authoring it

SERVED 2026-09-12 (hello-17): IssueRelation is the Linear rendering of canon's issue_link - the same row GitLab serves at /issues/{iid}/links. Linear names both ends outright (issue and relatedIssue) and keeps ONE word for the relationship, so Issue.relations is the rows an issue is the source of and Issue.inverseRelations the rows it is the target of, and neither needs GitLab's direction-dependent vocabulary. The type is mapped to Linear's own set - canon `relates` is Linear `related`, canon `duplicates` is `duplicate` - and the mapping THROWS on a kind with no Linear spelling rather than guessing; `parent` has no member in IssueRelationType at all and carries no issue_link row; parentage is Issue.parent and Issue.children, read from hello-23's own issue_parent table (clause added 2026-10-27, decisions/2026-10-27-2306-hello-23-serving-rest.md). IssueRelation.type is String! and not the enum, copied verbatim from the pinned SDL. On an artifact without the domain every relation field answers an explicit generation error in `errors` rather than an empty connection, which would assert that nothing in this universe is linked (invariant #4's GraphQL clause)

SERVED 2026-09-14 (founder batch A-R9: renderer work, no canon): a ProjectMilestone is canon's `milestone` seen through the epic whose issues carry it - the row gh. already publishes at /repos/{owner}/{repo}/milestones/{n} and gl. at /projects/{id}/milestones, now with a third rendering and nothing new asserted. THE PAIR IS THE ENTITY: Linear hangs a milestone off exactly one project (`ProjectMilestone.project` is non-null) while canon hangs it off a repository, so the checkpoint exists where issues carrying BOTH an epic and a milestone put it, and one canonical milestone reached through two projects is two checkpoints. `name` is `Sprint N`, the same string renderer-github's shapeMilestone prints for the same row, and BOTH instants come from the shared `ArtifactReader.milestoneTimes` accessor, so three hosts cannot date one milestone three ways (invariant #5). `progress` is the closed share of the issues the checkpoint actually holds, counted on the same two workflow states projectFacts counts, so a milestone and its project cannot disagree; `status` is measured (done when the milestone is closed or all its issues are, overdue when its date has passed with work open, unstarted when nothing has left the backlog, next otherwise). NO generation gate: `issue.epic` and `issue.milestone` are base-schema columns, so an older pin answers FEWER checkpoints - none at all in the hello-3 universe, where no issue carries both, and that empty IS the true answer there

refuse — Linear itself deprecated this surface: packages/conformance/specs/linear/graphql/schema.graphql:39381 marks `roadmap` and :39419 marks `roadmaps` with @deprecated(reason: `Roadmaps are deprecated, use initiatives instead.`), so mirroring it would advertise a surface Linear tells clients to stop calling (the ruled vendor-removed precedent, coverage/reviews/2026-09-01-gitlab-modes.md)