[kurze]

A small response goes out as a head write then a body write.
With Nagle's algorithm on, the body waits for the client's delayed ACK, about
40 ms on Linux (at least on my env), so every small object or manifest read on a kept-alive connection paid it.
A Go build cache over the DAV tree, mostly small entries, read back 3 to 4 times slower than with the option set.

useful if we want to use vk-registry as compilation cache server too

[Copilot]

🟡 Changes recommended

The ignored TCP_NODELAY result lacks the Rust-guideline-required explanation of why failure is safe to ignore.

1 open issue

• Explain why ignoring TCP_NODELAY failure is safe · New

What changed in this PR

This PR enables TCP_NODELAY on accepted registry connections to reduce delayed-ACK latency when clients repeatedly read small objects.

Changes:

• Set TCP_NODELAY in the registry’s connection acceptance loop.
• Add a regression test checking median latency for repeated small DAV reads.
• Document the latency improvement in the changelog.

File	Description
vk-registry/​tests/​dav_e2e.rs	Adds a small-object read latency regression test.
vk-registry/​src/​lib.rs	Enables TCP_NODELAY on accepted connections.
CHANGELOG.md	Records the small-read latency fix.

_{Review effort: Balanced}