v2.0.4

Improve cancellation diagnostics, accept videos with inaccurate dimension metadata, and harden deployments that share a single identity pool.

• Report caller cancellations as 499 warnings and shutdown cancellations as 503 service_shutting_down errors. Preserve cancellation through request-body parsing and Telegram response reads.
• Report interrupted response streams with their actual outcome, transferred byte count, and the status already sent as wire_status. Keep delivery outcome details consistent with shutdown attribution.
• Accept valid MP4s whose dimensions differ from upstream metadata. Buffered Telegram uploads use measured dimensions; streamed uploads omit automatic dimensions, and explicit caller values are preserved.
• Add persistent Dokploy settings for stop-first updates and rollbacks, readiness checks, a 60-second shutdown grace period, and failure monitoring. Add bounded Compose logs and deployment verification helpers.

Upgrade: allow 60 seconds for shutdown and run one active process per database schema. Follow the first-rollout procedure before applying the new Dokploy settings to an existing task: it retains its original shutdown limit. That procedure requires a maintenance window and briefly pauses scheduling on the single Swarm node. Host-wide Docker log rotation requires a separate Docker restart.

Validation: make check and make lint passed, including full race-enabled PostgreSQL tests, build, staticcheck, and govulncheck. All three PR CI checks passed. Deployment helper checks passed, and isolated Docker 29.7.2 Swarm tests verified readiness monitoring, graceful first rollout, lock release, recovery, and subsequent stop-first updates.

Pull request #21 · Full changes since v2.0.3