FE-1928: Pool atlas requests to handle network pressure better

[alex-e-leon]

🌟 Purpose

The current implementation of atlas fetches/prefetches all tiles naively, and doesn't handle network issues or network pressure well. If the server is overloaded, or the user is rate-limited the errors get forwarded to the client and a 429 message is shown. This PR introduces a network pool that slows down the network requests, especially pre-fetching if network errors are triggered.

In addition this PR also improves auth token refresh handling. In the original implementation if an auth token expires during network transit, only the first tile fetch reply renews the token and retries and any other tile fetches fail. We fix this by having all in flight tile fetches re-use and retry their requests with a new token during token renewal.

🚫 Blocked by

There is a seperate ticket to update the rate limits on atlas to better match a tiling workload:
https://linear.app/hash/issue/SRE-1170/adjust-atlas-rate-limits-for-tiling-workloads

I'd like for this to be finished first so that I can properly test the new logic against the updated rate limits.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
hash	Ready	Preview	Oct 13, 2026 3:49pm UTC

3 Skipped Deployments

Project	Deployment	Actions	Updated
hashdotdesign-tokens	Ignored	Preview	Oct 13, 2026 3:49pm UTC
petrinaut	Skipped		Oct 13, 2026 3:49pm UTC
petrinaut-docs	Skipped		Oct 13, 2026 3:49pm UTC

[cursor]

PR Summary

Medium Risk
Touches core atlas transport (retries, session replacement, concurrency) and tile caching behavior under load and auth renewal; mis-coordination could cause failed tiles or stale sessions, but changes are heavily tested.

Overview
Adds atlas request shaping so tile bursts do not overwhelm rate limits: a shared concurrency pool (cap of six) for data routes with required traffic ahead of low-priority prefetches, while manifest/current stay unpooled to avoid renewal deadlocks.

On 429 responses the transport now parses Retry-After, floors retry backoff on shorter waits, treats very long waits as non-retryable, and records speculation backpressure so prefetch pauses (and can cancel in-flight prefetches) until the budget clears.

Session recovery is reworked: withAtlasSession replaces the old replace-or-bail check so when a session ends, one caller replaces and concurrent siblings retry under the recorded successor instead of failing for acting second.

Tile cache / prefetch changes: auxiliary detail is retained with tiles (labels count toward budget), prefetches run at the viewport’s detail mode, and getViewportNodes skips speculation while backpressure is active.

^{Reviewed by Cursor Bugbot for commit 9aafd93. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

There are 2 total unresolved issues (including 1 from previous review).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit b1b1569. Configure here.}