Use short-lived detection execution transactions

[fengting124]

What

• split detection claim, external inference, and result persistence into short transaction boundaries
• add execution leases and fencing tokens so stale attempts cannot overwrite newer work
• keep multi-model results atomic and leave busy Redis messages pending for safe redelivery
• document transaction, recovery, and operational semantics

Why

The previous implementation held a database transaction for the entire model HTTP call. Slow inference consumed a database connection and a crashed worker had no durable execution ownership. This change makes model latency independent of database transaction duration while preserving recoverability.

Guarantees

• only the current execution token may persist success or failure
• expired attempts can be replaced; late responses are discarded
• no partial predictions are stored when one model in an ensemble fails
• model weights and GPU runtime remain outside this change

Verification

• Java: 97 tests passed
• frontend: 8 tests, lint, and production build passed
• model service: 6 tests passed without weights
• smoke tools: 3 tests passed
• integration test asserts model inference runs without an active Spring transaction