#36259 http: update m_keep_alive under m_send_mutex

Synchronizes `HTTPRemoteClient::m_keep_alive` updates with `m_send_buffer` by moving the assignment inside `m_send_mutex` and switching the field from `std::atomic_bool` to `bool GUARDED_BY(m_send_mutex)`. Also adds a regression unit test simulating the race condition.

Problem: When a client pipelines a keep-alive request followed by a 'Connection: close' request, the worker handling the second request clears `m_keep_alive` before taking `m_send_mutex`. If the I/O thread drains the first response in between, it observes an empty buffer with keep-alive disabled and disconnects the client before the second response can be queued, dropping it.

Ready

Nothing known that would invalidate a review now.

Waiting on (from the thread read):

• author: response to winterrdog's suggestion regarding pragmatism and priority 2026-09-15

Strong

substantive support, no open objection (winterrdog)

Thread notes: Added pinheadmz's non-substantive Concept ACK.

Objections, suggestions, and questions:

Support:

• winterrdog (contributor) (Concept ACK, 2026-09-15) [read]: makes sense to update m_keep_alive in the same critical section as the send buffer, since both are read together under m_send_mutex, keeping state changes in sync and preventing the I/O thread from seeing an empty buffer with keep-alive cleared
• pinheadmz (member) (Concept ACK, 2026-10-31): seems reasonable, but low priority for review [not substantive]

Participants:

• winterrdog (contributor, support, 1 statements 2026-09-15 to 2026-09-15): Concept ACK, but suggested delaying review because the scenario is hard to reach
• pinheadmz (member, support, 1 statements 2026-10-31 to 2026-10-31): Concept ACK, considered low priority for review

144 lines under test/bench/ci.

• src/test/httpserver_tests.cpp +144/-0
• src/httpserver.cpp +8/-2
• src/httpserver.h +4/-1

The short text the ranking pass compares across PRs.

This PR fixes a race condition in the internal HTTP server where HTTPRemoteClient::m_keep_alive was updated outside m_send_mutex. For pipelined requests where a keep-alive request is followed by a connection-closing request, the I/O thread could observe an empty buffer and a disabled keep-alive flag between requests, disconnecting the client and dropping the second reply. The fix guards m_keep_alive with m_send_mutex and includes a deterministic mock-socket regression test. winterrdog provided a concept ACK supporting the lock synchronization, and the PR is ready for review.

P3 · bug fix (ranking pass: Moved up 13: dropped HTTP responses on pipelined requests is a genuine server-side correctness race, ahead of feature additions.)

Fixes an HTTP server race condition where pipelining a keep-alive request followed by a 'Connection: close' request can cause the I/O thread to disconnect prematurely and drop the second response. While a legitimate correctness fix for pipelined HTTP requests, it addresses a narrow edge case that is reasonably deferrable without broader impact on RPC stability.

Membership: Matches the RPC/REST/ZMQ label and directly modifies the HTTP server in src/httpserver.cpp and src/httpserver.h.

Factors: security/stability 1, bug 1, performance 0, user value 1, leverage 0