fix: make local docker uploads replica-independent (#104)

## Why

Chunked blob uploads kept the in-progress session in **process memory** keyed by upload UUID, so the `POST`/`PATCH`/`PUT` of a single `docker push` had to land on the same replica. The API runs at `minReplicas: 2` with no session affinity (see argocd-apps `api-hpa.yaml`), so a real push — which streams the layer via `PATCH` then finalises with `PUT` — intermittently 404s with `BLOB_UPLOAD_UNKNOWN` when a chunk hits a replica that never saw the `POST`. This was flagged when the local docker registry landed (#103).

## Changes

- Stage chunked uploads in object storage under `uploads/<uuid>` instead of an in-memory temp file. The UUID travels in the `Location` URL handed to the client, so any replica reconstructs the staging key with no shared in-process state. Finalise streams the staged bytes plus any trailing `PUT` body through the CAS in one pass; monolithic uploads are unchanged.
- Support `DELETE` of an in-progress upload (cancel) by dropping its staging object.
- Reap abandoned staging objects in the GC (`uploads/` older than 24h) via a new `S3.ListStaleObjects`, so cancelled/interrupted pushes don't leak.

## Verification

- Split a single push across **two instances sharing one Postgres+MinIO**: `POST`→A, `PATCH`→B, `PUT`→A finalises with the correct digest, and the blob pulls back **byte-identical from both** replicas. Config-blob and manifest pushes split the same way succeed; `tags/list` is correct. (Pre-fix, the cross-replica `PATCH` 404s.)
- `scripts/docker-e2e.sh` still passes (incl. `TestLocalDockerPushPull`); unit tests + `go vet` clean.

Reviewed-on: #104
Co-authored-by: Ben Vincent <ben@unkin.net>
Co-committed-by: Ben Vincent <ben@unkin.net>

internal/gc/gc.go

@@ -14,6 +14,11 @@ import (
 // before the referencing artifact/local_files row exists.
 const blobGracePeriod = 1 * time.Hour
 
+// uploadGracePeriod is how long a docker blob-upload staging object
+// (uploads/<uuid>) may sit idle before GC treats it as an abandoned push and
+// reaps it. Generous so a slow but live push is never cut off mid-flight.
+const uploadGracePeriod = 24 * time.Hour
+
 type Collector struct {
 	db       *database.DB
 	store    *storage.S3
@@ -43,6 +48,8 @@ func (c *Collector) Run(ctx context.Context) {
 func (c *Collector) sweep(ctx context.Context) {
 	start := time.Now()
 
+	c.sweepUploads(ctx)
+
 	orphaned, err := c.db.FindOrphanedBlobs(ctx, blobGracePeriod)
 	if err != nil {
 		slog.Error("gc: find orphaned blobs", "error", err)
@@ -70,3 +77,24 @@ func (c *Collector) sweep(ctx context.Context) {
 		)
 	}
 }
+
+// sweepUploads reaps docker blob-upload staging objects abandoned longer than
+// uploadGracePeriod (cancelled or interrupted pushes that never finalised).
+func (c *Collector) sweepUploads(ctx context.Context) {
+	stale, err := c.store.ListStaleObjects(ctx, "uploads/", time.Now().Add(-uploadGracePeriod))
+	if err != nil {
+		slog.Error("gc: list stale uploads", "error", err)
+		return
+	}
+	reaped := 0
+	for _, key := range stale {
+		if err := c.store.Delete(ctx, key); err != nil {
+			slog.Warn("gc: delete stale upload", "key", key, "error", err)
+			continue
+		}
+		reaped++
+	}
+	if reaped > 0 {
+		slog.Info("gc: reaped stale docker uploads", "count", reaped)
+	}
+}