Optimize Spark PDF translation throughput

[SunChJ]

Summary

• replace the fixed 12-item BabelDOC cap with an estimated character budget
• use a 3,000-character default selected from repeated local Spark measurements
• recover malformed structured output by bisecting only the failed document batch
• pace Spark background turns and retry capacity errors without applying the policy to other models
• add an opt-in reproducible PDF package benchmark

Local benchmark

18,006 source characters from 2603.05890v1.pdf, Spark with low reasoning:

• 3,000 chars: 4/4 successful, 51.3 s median, 12 turns median
• 2,600 chars: 4/4 successful, 63.1 s median, 14 turns median

Validation

• swift test --filter LoopbackServerTests (9 passed)
• swift test --filter CodexAppServerClientTests (24 passed)
• swift test --filter TranslationDispatchCenterTests (5 passed)
• live Spark package benchmark (11 measured runs across screening and finalists)