perf(relay): cache Git pack hydration

[thomaspblock]

Summary

Git clone, fetch, and push hydration now reuses verified immutable pack/index pairs instead of downloading and indexing unchanged packs for every request. Object storage remains authoritative; cache misses, eviction, and relay restarts affect performance only.

The cache is content-addressed by pack digest, coalesces concurrent misses into one population flight, and hard-links cached files into each isolated request workspace with a copy fallback. Retained bytes, population concurrency, pack/index sizes, stale sessions, and the per-pod cache volume are all bounded, with Prometheus metrics for hits, misses, coalescing, population waits, size, activity, eviction, and fallback behavior.

Operational design

• Each relay process owns an isolated cache session on a per-pod ephemeral volume.
• A heartbeat removes abandoned sessions after crashes without touching active sessions.
• The default 5 GiB retained cache allows at most two concurrent cold populations on a 7 GiB volume.
• LRU eviction never removes files still being linked into an active request workspace.
• Repository, owner, community, and digest values are excluded from metric labels.

Test plan

• cargo clippy -p buzz-relay --all-targets -- -D warnings
• cargo test -p buzz-relay --lib api::git::pack_cache::tests
• cargo test -p buzz-relay --lib api::git::hydrate::tests
• cargo test -p buzz-relay --lib api::git::transport::track_c_tests
• cargo test -p buzz-test-client --test e2e_git -- --ignored --nocapture
  • Verified clone, push, fetch, force-push, concurrent push rejection, and repository recovery against a local relay and MinIO.
  • The smoke run recorded 46 cache hits, 6 misses, 6 retained entries, and zero active populations after completion.