fix(storage): make GCS uploads resilient to network/DNS stalls

[raaymax]

Problem

Prod file uploads (chat.codecat.io) intermittently fail: the progress bar fills but never turns green, then the message send fails. It came and went over whole days and self-healed, with no app-level error logs.

Root cause

Uploads are streamed to GCS during the request, and the GCS client made 2–3 sequential outbound calls to Google (auth token → object POST → metadata PATCH) with no timeout and no retry. Any transient stall in the container's egress/DNS path (single upstream resolver via Docker's embedded DNS, plus parallel A/AAAA lookups) made those fetch calls hang indefinitely — hanging the whole upload with nothing logged.

Amplifier: the deployed main cached the GCS access token for only 2 seconds, so nearly every upload re-authenticated against oauth2.googleapis.com, multiplying the round-trips that could stall. (dev already caches for 50 min; this PR keeps that and hardens the rest.)

Changes

• gcs.ts — every GCS request (auth token, upload POST, metadata PATCH, stat, list, get, remove, exists) now goes through a gcsFetch wrapper with a per-attempt AbortSignal.timeout + bounded retries + exponential backoff. A stalled path aborts and retries instead of hanging; transient 429/5xx are retried.
  • Upload payloads within a size limit are buffered so a failed attempt can be safely replayed; larger files stream in a single attempt (no unbounded memory).
• http/mod.ts — re-enabled concise HTTP access logging (method url status latency), escalated to warn on 5xx or >=2s, so upload issues are visible instead of silent.

Configurable via STORAGE_GCS_TIMEOUT_MS, STORAGE_GCS_UPLOAD_TIMEOUT_MS, STORAGE_GCS_MAX_RETRIES, STORAGE_GCS_BUFFER_LIMIT_BYTES.

Testing

• deno fmt / deno lint / deno check clean on both files.
• Verified end-to-end against the real GCS bucket from inside the prod container: upload 200, stat, get (bytes match), remove, exists-after-remove — all correct via the new buffered/retry path.

Deployment note

The deployed main also carries the 2-second token-cache bug (fixed on dev), so shipping this via dev → main also delivers that fix. For immediate prod relief, this commit can be cherry-picked onto a hotfix branch off main.

Related infra (out of scope for this PR)

Host-level IPv6 was misconfigured (rogue ULA RAs from smart-home/RPi devices); the servers now ignore RAs (accept_ra=0). That's hygiene — this code change is what makes uploads resilient regardless of the network path.