Greptile Summary

This PR replaces the per-downloader transport construction (and the old per-config sync.Map cache in the single downloader) with a centralized NetworkPool that ref-counts leases and evicts idle transports after 10 seconds. Bootstrap and download clients in the concurrent downloader now share the same pool entry (both use PoolMaxConnsPerHost), resolving the previous key-mismatch concern. The probe layer drops its LRU client cache and acquires from the same pool per call.

Confidence Score: 5/5

Safe to merge; the one remaining finding is a minor test consistency issue that does not affect correctness.

All previously raised P0/P1 concerns (IPv4 preference, test race, bootstrap/download key mismatch) are addressed or intentionally resolved. The only new finding is a P2 pool-key inconsistency in the prewarm test that doesn't affect production behavior.

internal/engine/concurrent/prewarm_reuse_test.go — pool key should match the one used in Download().

Important Files Changed

This is a comment left during a code review.
Path: internal/engine/concurrent/prewarm_reuse_test.go
Line: 27-29

Comment:
**Pool key mismatch between test and production code**

`AcquireTransport` is called here with `runtime.GetMaxConnectionsPerHost()` (64 by default), but `Download()` always passes `types.PoolMaxConnsPerHost` (512). These resolve to different `poolKey` entries, so the transport acquired in the test has `MaxConnsPerHost = 64` while the one used in production has `MaxConnsPerHost = 512`. The prewarm→reuse assertion still passes because both use the same client object, but the test is no longer exercising the transport configuration that actually runs in production. Consider passing `types.PoolMaxConnsPerHost` here to stay in sync with `Download()`.

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (5): Last reviewed commit: "refactor: standardize network transport ..." | Re-trigger Greptile}