fix: 0.17.2 — sqs_consumer secrets wait, FIFO packaging, canvas updating UX

[drernie]

Summary

Three independent user-visible fixes landing as 0.17.2.

1. sqs_consumer waits for Benchling secrets instead of crashing (ddf648f)

Addresses the 🔴 "CI deploy_dev_stack failing" blocker sir-sigurd flagged on quiltdata/deployment#2357. Root cause lives here in the image, not in the deployment template.

Before: sqs_consumer.main() called config.get_benchling_secrets() unconditionally at startup. On a fresh deploy where BenchlingSecret is created empty by design (populated later by the external config script), the fetch raised SecretsManagerError → process exited → Essential=True cascaded → ECS Deployment Circuit Breaker tripped → UPDATE_ROLLBACK_COMPLETE.

The HTTP container is immune because it records the failure via _runtime_guard_response and keeps serving 503s. The sidecar had no equivalent safety net.

• Pre-start readiness wait (docker/src/sqs_consumer.py). Signal handlers installed before wait_for_ready_config() loops with bounded exponential backoff (30s → 300s) until get_benchling_secrets() + apply_benchling_secrets() succeed. SIGTERM short-circuits the wait via asyncio.Event so ECS can stop the container cleanly even during the wait. Catches both SecretsManagerError (secret empty/malformed) and ValueError (env var missing).
• MaxNumberOfMessages = min(10, concurrency) — default concurrency is 5; batching up to 10 would have 6–10th messages sitting in the semaphore backlog eating visibility timeout.

2. Canvas shows full layout with "Updating..." during export (beb3b3c)

Replaces the bare "Processing..." placeholder canvas with the full navigation layout (Browse/Update buttons + package header + footer), where the footer reads "Updating..." until the background workflow delivers the real "Updated at" timestamp. Update Package is disabled while updating to prevent a second concurrent export; Browse Package stays enabled so the previous package version remains reachable on re-exports. Skips the linked-packages Athena query on the initial render to keep it fast on the request thread.

3. FIFO SQS per-entry sequencing replaces .canvas_id sidecar (4e31324)

Webhook handlers used to fire-and-forget by spawning daemon threads, which let entry.created and canvas.created events for the same entry race to write entry.json. The .canvas_id S3 sidecar masked one symptom of that race (canvas_id loss), but the deeper issue was no per-entry sequencing.

Replace the daemon-thread "queue" with a real FIFO SQS queue keyed by MessageGroupId=entry_id. A new BenchlingPackagingConsumer sidecar drains the queue and runs EntryPackager.execute_workflow. Sequential processing per entry makes the sidecar unnecessary; a late entry event simply reads canvas_id from any existing entry.json.

• Add PackagingRequestQueue.fifo (+ DLQ) in CDK with 40-min visibility timeout and content-based dedup
• New packaging_publisher module: webhook handlers enqueue here
• New packaging_consumer module + Fargate sidecar that drains the queue
• Refactor SqsConsumer into a BaseSqsConsumer + two subclasses
• Delete _save_canvas_id, _load_canvas_id, _load_canvas_id_sidecar, and execute_workflow_async; existing .canvas_id files in S3 become harmless orphans
• Move synchronous send_updating_canvas into the canvas webhook handler

Test plan

• npm test — lint + typecheck + TS + Python unit tests
• npm run version:tag:dev triggers CI image build
• npm run test:dev against deployed image
• Verify on quiltdata/deployment#2357 that deploy_dev_stack no longer trips circuit breaker on fresh deploy
• Manual canvas export: confirm "Updating..." footer with Update disabled / Browse enabled
• Manual concurrent entry.created + canvas.created replay: confirm canvas_id preserved in entry.json

🤖 Generated with Claude Code