Backstage, read from its own source and never installed: a scaffolder that runs once against a golden path that reconciles forever

Artifact — an answer, cited from Backstage's own source on forge, to one structural question: its scaffolder is a one-shot linear task runner — no reconcile loop, no drift detection, no desired-state comparison — the opposite of everything Group A taught. Backstage is read, not installed: its yarn tsc build OOMs at ~4GB and no --set fixes it, so the disqualifier is the build, not the runtime — which is itself the lesson about why it does not fit here.

Rests on — the golden path you built, the Flux-reconciled CRD that this reads against; and the drift beat, because "runs once" versus "converges continuously" is the exact axis this compares on.

Topology — no cluster. This is a source clone on forge; platform may be up and idle beside it or torn down — the reading does not touch it.

Read — the module framing and the question below, against Backstage's actual scaffolder code. Held to the archaeology drill standard: the "runs once" claim resolves to a function a reader can open, not to a blog assertion.

Question to answer (from the tools against each other): a Backstage scaffolder template and a Flux-reconciled CRD both "create an app from a form." One converges continuously and one runs once. Which failure does each hide from the developer, and which does it expose? Name a concrete case where the difference bites — e.g. someone hand-edits the generated Deployment: the reconciled CRD corrects it, the scaffolder never sees it was created.

Do — clone Backstage on forge and read the scaffolder, without building it:

ssh zain@forge 'git clone --depth 1 https://github.com/backstage/backstage && \
  ls backstage/plugins/scaffolder-backend/src/scaffolder'
# read the task runner: it executes a list of steps once and stops. Find where.
# DO NOT run `yarn install && yarn tsc` — it scaffolds NODE_OPTIONS=--max-old-space-size=8192
# and still OOMs at ~4GB. The build is the disqualifier, not the runtime.

Verify from outside — the "one-shot" claim is a citation: a reader opens the scaffolder's task-execution path at your stated commit and finds a linear run of steps with no requeue, no watch, no resourceVersion — structurally unlike the P4 loop every Group A tool is. "Backstage doesn't reconcile" fails the gate; <file>:<function>@<commit> passes.

Expect — a precise structural contrast, not a preference: the scaffolder is strictly more interesting to compare than to install, because its lack of a control loop is the whole point next to Flagger and Flux in the same phase. The golden path you built from a reconciled CRD does something Backstage's scaffolder cannot — notice drift.

Write down — the cited scaffolder function proving "runs once," and the concrete case where one-shot-versus-reconciled bites a developer. This feeds the failure-mode catalogue and the capstone critique.

Footprint note — a source clone on forge, which is never part of any topology; do not attempt the Backstage build anywhere — it exceeds the ceiling on its own.

Teardown — remove the clone from forge. No cluster state was touched. Whatever topology was idle beside it stays.