fix(race+throughput): close 3 throughput-limiting gaps

[zbnerd]

Summary

3 independent fixes that together restore item-equipment throughput to 150 files/s (from 102). All verified end-to-end on 2026-06-16, heap 2g.

1. ext-api/ChunkedSnapshotSink: chunk-ready publish race

• Use S3TransferManager future.whenComplete (not blocking join)
• Order preserved: publish only fires after PUT completes successfully
• Writer thread returns in ~50ms (gzip+close), not 1-4s
• 0 race errors vs prior 1133 failed chunks/90s

2. ext-api/application.yml: in-flight 100 → 250

• 250-permit rate limiter was being throttled by 100-concurrent batch cap. Aligning both unblocks the rate limiter.
• Per-batch wave time 2.5s → 1.6s (with heap fix in PR chore(ext-api): bump heap -Xmx1g to -Xmx2g #1293)

3. infra/MinioObjectStorage: drop temp-file double-spool

• Was: Files.createTempFile + Files.copy + putObject + delete = 4 round-trips/chunk, /tmp I/O contention
• Now: drain stream to ByteArray, RequestBody.fromByteArray = 1 round-trip, no disk
• Sync S3Client.putObject cannot chunked-stream (no length=-1 API on sync path), so in-memory is the only sync option

4. calculator/CurrentRunIdHolder: in-memory set → polled from ext-api /run-status

• Was: ConcurrentHashMap (lost on restart, drift on multi-instance)
• Now: polled from ext-api /run-status endpoint
• Coordinator + test refactored to match
• 4 files in module-calculator

Test plan

• ext-api item-equipment: 102 → 150 files/s
• calc downstream: 186 → 362 users/s
• 0 race errors after 100+ chunks
• calculator skip reason 'stale_run' replaces 'endpoint_mismatch'
• 4 modules health UP, MinIO health UP

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.