[mcp] Streaming response for batch jobs > 1000 emails

[Davison-Francis]

Currently deliveriq_batch_download returns the full CSV in one tool result. For large jobs this either blows up the LLM context or truncates silently.

Proposed: switch to a paginated tool — deliveriq_batch_download_page(jobId, cursor) — that yields ~500 rows per call, with a continuation cursor in the response. Claude can then loop until the cursor is null.

Why this matters: the most useful batch flow is 'verify 5,000 emails from this CSV, then summarize the bad ones' — currently this fails silently when the CSV is too big.

Happy to take a PR.

[Davison-Francis]

Roadmap status:

The current deliveriq_batch_verify already supports async submit + poll + download (separate _status and _download tools), so for batches >1000 the recommended pattern is:

1. Submit via deliveriq_batch_verify (returns a job_id immediately, no streaming)
2. Poll deliveriq_batch_status with the job_id (returns per-category breakdown as it progresses)
3. Download with deliveriq_batch_download once status == "complete"

For true streaming via MCP itself: that's blocked by the current MCP spec — tools/call returns a single response, not a stream of partial results. When the spec adds streaming tool responses (work-in-progress at modelcontextprotocol/specification), we'll add a deliveriq_batch_stream tool that yields per-record results as they complete.

Workaround in the meantime: poll deliveriq_batch_status more aggressively (every 5s instead of every 30s) and use the per-category counts to show a progress bar.

Targeted for the v1.2.x line once MCP streaming lands.