GC engine (internal/gc/): Collector.Run() implements the two-phase
algorithm — Phase 1 finds unreferenced blobs and deletes DB rows in
a single transaction, Phase 2 deletes blob files from storage.
Registry-wide mutex blocks concurrent GC runs. Collector.Reconcile()
scans filesystem for orphaned files with no DB row (crash recovery).
Wired into admin_gc.go: POST /v1/gc now launches the real collector
in a goroutine with gc_started/gc_completed audit events.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Phase 2 — internal/storage/:
Content-addressed blob storage with atomic writes via rename.
BlobWriter stages data in uploads dir with running SHA-256 hash,
commits by verifying digest then renaming to layers/sha256/<prefix>/<hex>.
Reader provides Open, Stat, Delete, Exists with digest validation.
Phase 3 — internal/auth/ + internal/server/:
MCIAS client with Login and ValidateToken, 30s SHA-256-keyed cache
with lazy eviction and injectable clock for testing. TLS 1.3 minimum
with optional custom CA cert.
Chi router with RequireAuth middleware (Bearer token extraction,
WWW-Authenticate header, OCI error format), token endpoint (Basic
auth → bearer exchange via MCIAS), and /v2/ version check handler.
52 tests passing (14 storage + 9 auth + 9 server + 20 existing).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
