A proven-commit import applies modifications but not additions — the partition is left referenced-but-incomplete, and a force update imports 0 nodes #4456

What is failing

A GitSync import at a proven commit applies modifications to a partition but does not create files the same commit adds. The partition is then referenced-but-incomplete, its NodeTypes cannot compile, and the bake gate holds readiness — so the instance cannot roll. A forced re-import does not heal it: it imports 0 nodes.

Measured on memex.systemorph.com, 2026-09-15, rolling onto sealed set 3.0.0-ci.8692 (core 3cd1935709, plugins sealed at 85515f62).

The import

Hosting/_Activity/10f16763 — "Update Hosting to the built commit 85515f62", 19:00:52 → 19:01:48:

Recompiling 8 NodeType(s): Hosting/Admin, Hosting/Deployment, Hosting/Backup, Hosting/FleetConsole,
                           Hosting/InstanceAction, Hosting/InstanceRequest, Hosting/Issue, Hosting/PlatformBuildInbox
Adopted 8 prebuilt assembly(ies) — the rest compile.
Pruned 118 node(s) absent from the repo: Hosting/Triage/ci-failure/…, Hosting/Triage/ci-green/…, Hosting/Triage/Status
Re-imported ImportedWithErrors (41 node(s)) at 85515f62.

Hosting/_GitSync afterwards: lastAttemptedCommitSha: 85515f62, lastSyncOutcome: ImportedWithErrors, lastSyncCommitSha still c7f03952.

What landed and what did not

In Hosting/Deployment/Source/, that import updated two existing nodes — PlatformBuildInboxWatcher (v14) and HelmValues (v8), both lastModified 2026-09-15T19:01:12Z. In the same commit range it did not create Hosting/Deployment/Source/SelfUpdateRouting. The partition holds 13 Code nodes in that folder and it is not among them.

SelfUpdateRouting.cs is present at 85515f62 (git ls-tree), and at that same commit it is referenced by Hosting/Deployment/Source/PlatformBuildInboxWatcher.cs — the very file the import did apply — and by Hosting/PlatformBuildInbox/Test/PlatformBuildInboxTests.cs. So the import wrote the reference and skipped the referent.

🚨 Hosting/ is byte-identical between 85515f62 and today's main (git diff --stat 85515f62 origin/main -- Hosting is empty). This is not a "newer content" or seal-ordering problem: the importer had one consistent tree and applied it partially.

The consequence

On the 3.0.0-ci.8692 replica (memex-portal-deployment-6546c779d4-4ngqv):

CS0103 Error (line 6362): The name 'SelfUpdateRouting' does not exist in the current context
CS0246 Error (line 7308): The type or namespace name 'SelfUpdateRouting' could not be found
DynamicTypePreWarmer: REFUSING READINESS — 5 NodeType(s) regressed on this image:
  Essentials/Mailbox, Hosting/InstanceAction, Hosting/Issue, Hosting/ModuleReportInbox, Hosting/PlatformBuildInbox
BuildProtocol: NOT publishing GO for s4b2836629f0c58ad4ed0aa683190981a — 5 gating regression(s)

The gate behaved correctly: the previous replicas kept serving and there was no outage. But the rollout cannot complete, and the helm-release deploy failed at "Observe the rollout until it completes" after 28 minutes.

The repair path does not repair

git_hub_sync update force: true on Hosting (branch HEAD = main, whose Hosting tree is identical to the sealed commit) — Hosting/_Activity/89ae5941:

Pruned 6 node(s) absent from the repo: Hosting/Triage/…
Imported ImportedWithContentErrors (0 node(s)).

Zero nodes imported, and the missing file is still missing. A force update is documented as "overwriting local changes"; it cannot create the absent node either. That is what makes this a hard stop rather than an inconvenience: there is no supported operation that restores the partition.

Same shape, second partition

Essentials/Mailbox fails identically on MailboxLiveTests (CS0103/CS0246, 7 sites): the referencing test-area file is in the mesh, Essentials/Mailbox/Test/MailboxLiveTests.cs is not. Unlike Hosting, Essentials/ on main is 10 files ahead of 85515f62, so that partition additionally carries content newer than the seal — the two-writer mix of #4355 — but the immediate failure is the same missing-addition shape.

Why this is worth fixing as its own defect

#4355 is about two writers disagreeing over one partition. This is narrower and, I think, more serious: one importer, one consistent tree, and additions silently dropped — with ImportedWithErrors / ImportedWithContentErrors as the only signal, and no per-node error in the activity log naming what failed. The outcome string is on the sync config where nothing reads it; the activity's maxSeverity is Information and its status is Succeeded.

Suggested, in order of value:

1. An import that could not write a node must say which node and why, at Warning or Error, in the activity log — ImportedWithErrors (41 node(s)) names neither the failures nor the count of them.
2. A force update must be able to create a missing node. Today it imported 0 while a referenced file was absent from the partition.
3. A partition left referenced-but-incomplete should not be handed to the compiler, or the compile failure should name the import that produced it — the operator's visible symptom is CS0103 on a symbol whose file is in git.

Evidence

• Hosting/_Activity/10f16763, Hosting/_Activity/89ae5941, Hosting/_GitSync, Essentials/_GitSync on memex.systemorph.com
• pod memex-portal-deployment-6546c779d4-4ngqv (3.0.0-ci.8692), 18:56Z onwards
• Memex deploy run 35010439754 (failed at Observe the rollout), Ops/Actions/helm-memex-20260915-deploy-8692
• git ls-tree -r 85515f62 -- Hosting/Deployment/Source/SelfUpdateRouting.cs; git diff --stat 85515f62 origin/main -- Hosting (empty)