Fix transient memory amplification in batch_get_documents_tool

[mattpodwysocki]

Summary

The previous cache fix bounded persistent memory growth but left a transient amplification path open: a batch of 20 URLs that all normalize to the same page would still trigger 20 concurrent HTTP fetches and buffer 20 response bodies before the cache could help.

• Intra-batch URL deduplication: before any fetches are started, URLs are grouped by their normalized key. All inputs that resolve to the same page share a single in-flight Promise — so 20 cache-busting variants of one URL result in exactly 1 HTTP request, not 20.
• Streaming body size cap (readBodyWithLimit): checks Content-Length upfront and rejects before reading; if Content-Length is absent, aborts the stream as soon as accumulated bytes exceed the 2 MB limit. Avoids buffering oversized responses entirely.
• Same body size cap applied to get_document_tool.

Test plan

• 6 new tests in test/tools/batch-get-documents-tool/BatchGetDocumentsTool.test.ts
  • 20 cache-busting URLs → 1 HTTP call, all results return same content
  • 2 distinct normalized URLs → 2 HTTP calls
  • Already-cached URL → 0 HTTP calls
  • Content-Length over limit → per-URL error, batch continues
  • Non-mapbox URLs rejected before any fetch
  • Non-ok HTTP response → per-URL error
• All 40 existing tests still pass
• CI green

🤖 Generated with Claude Code