improvement(search): shorter Jira any-term budget and a shared batched read budget - #9984

Summary

• Jira's any-term fallback gets 2s instead of 4s. Successful fallbacks finish within ~1.5s (p95; p99 2.5s), while about a third of attempts stall until the cut, and the whole federated search waits on that stall
• Scaled MCP reads return up to 20K characters per document whatever limit the client sends (a live document is one chunk, so the chunk limit never sized it). Clients almost never raised limit, so they got 8K pages and came back for continuations
• Batched reads share one 40K-character call budget: short documents keep all their text and leave the rest to long ones, which continue through next. No document drops below the standard 8K page, so large batches read what they read today
• tools/list is unchanged, so clients aren't asked to re-approve the connector

Type of Change

• Improvement

Testing

• bun run test lib/sim-search lib/knowledge/mcp (527 passed); the new read-budget.test.ts, read-window and fallback-budget tests fail against the previous code
• biome on changed files
• A/B on our Search MCP benchmark (105-case bundle, 2 control and 2 candidate arms, 2 attempts each): no case flipped from right to wrong; correctness 71.5% → 74.4% (CI spans 0); completeness and source recall flat; rounds −0.13, tool calls −0.20, searches −0.15 per question (CIs above 0)

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing (new tests pass the test-audit authoring gate)
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)