Retain complete paper-batch inputs before final projection

[witbrock]

What changed

• declare the final batch projector's mandatory input keys as represented durable reads
• preserve message_iteration_results and message_iteration_errors losslessly across the checkpoint before batch_result is constructed
• retain optional pagination fields as optional
• bump the email-source workflow seed to version 17
• strengthen the checkpoint regression to preserve 100 rows and nested specific error detail under a 512-byte total budget

Root cause

Seed v16 preserved batch_result once it existed, but the source arrays were compacted immediately after process_messages. The next state therefore built an already-truncated batch result: at most eight item rows and depth summaries instead of failure causes.

Live evidence

Instance 00072b84-fd69-40b3-9b56-3bcc6d34a9c9 completed as done_with_retryable_failures with 10 selected, 8 review dispositions, and 2 failures. A same-conversation follow-up found the retained parent under a 27-to-20 history projection and accounted for all messages, but explicitly reported the item-array and nested-error truncation. This PR fixes that boundary for the larger 22-item general case.

Validation

• 311 affected-path tests passed
• 100 item rows and nested error details survive a forced 512-byte checkpoint budget
• no optional-pagination regression

Jira: JVNAUTOSCI-2635