#803 Datastore should fail fast on unresolvable credentials instead of stalling on the AWS provider chain

When @swamp/s3-datastore runs with no resolvable AWS credentials (no AWS_PROFILE, no default profile, no static env creds), the AWS SDK provider chain doesn't fail quickly — it falls through SSO token resolution to the EC2 IMDS endpoint (169.254.169.254). Off-EC2, that's either a fast-fail (~2s, when an expired SSO token short-circuits it) or a long stall (tens of seconds, IMDS connect-timeout × the sync layer's 30s request timeout × 3× retry backoff).

Measured, same machine back-to-back: profile pinned → datastore status healthy in ~22ms; no profile → CredentialsProviderError after 2011ms (fast path), with much longer stalls observed when no token is cached.

The stall is the harmful case: a push/sync that overruns a caller's command timeout gets SIGKILL'd mid-hold, which orphans the datastore lock (the #218 / #666 orphan path) and degrades concurrency for everyone.

Proposed: a short, bounded credential preflight before any S3 op (e.g. ≤2–3s for the whole chain) that aborts with an actionable error if creds don't resolve; and, when not on EC2, skip or sub-second-cap the IMDS provider (e.g. set AWS_EC2_METADATA_DISABLED internally) so off-EC2 hosts never eat the multi-minute hang. Fast, loud failure means the process exits before a timeout-SIGKILL can orphan a lock.

Relates to #666 (lock contention) and #218 (orphan loop) — this removes one trigger that feeds both.

Upstream repository: https://github.com/systeminit/swamp-extensions

Environment

• Extension: @swamp/s3-datastore@2026.06.03.2
• swamp: 20260617.212026.0-sha.396e0952
• OS: darwin (aarch64)
• Deno: 2.8.3
• Shell: /bin/zsh