[BUG] An SSE fault-stream slot is held for up to 30 s after the client disconnects

[bburda]

Bug report

Steps to reproduce

1. Start a gateway with the default sse.max_clients (2).
2. From one client, connect to GET /api/v1/faults/stream, disconnect, and reconnect, about once
   per second.

Expected behavior

A slot is freed when the connection ends, so a single reconnecting client cannot lock itself out.
The default should also be higher than 2, since a browser, a recorder and an operator may all
want this stream at the same time.

Actual behavior

After a few cycles, from that same single client:

HTTP 503
{"error_code": "service-unavailable",
 "message": "Maximum number of SSE clients reached. Please try again later."}

Gateway start-up line:
SSE fault handler initialized, subscribed to /fault_manager/events, max_clients=2.

One client in a reconnect loop blocks the stream for everyone, including the Web UI.

Environment

• ros2_medkit version: main (a268379), gateway vendor_info.version 0.6.0, SOVD API 1.0.0
• ROS 2 distro: Jazzy
• OS: Ubuntu 24.04

Additional information

The release itself is correct. The release_guard deleter
(src/http/handlers/sse_fault_handler.cpp:375-389) is destroyed together with the stack-local
httplib::Response. The problem is when that happens.

cpp-httplib checks strm.is_writable() only at the top of each write_content_chunked
iteration, which means only after the content provider returns true
(src/vendored/cpp_httplib/httplib.h:4206-4217). The fault-stream provider never returns true.
src/http/handlers/sse_fault_handler.cpp:310-349 is an inner while (true) whose only exits are
return false, and the return true at :351 cannot be reached.

So a closed connection is only noticed when the next write happens, either an event flush
(:285) or a keepalive (:333). Between events the loop waits in queue_cv_.wait_for (:319)
for keepalive_interval_, which defaults to kKeepaliveIntervalSec = 30
(include/ros2_medkit_gateway/http/handlers/sse_fault_handler.hpp:290). There is no ROS
parameter for it. A shorter value is only reachable through the test-only constructor.

On a quiet system the window is the full 30 s.

The other two SSE producers do return true on each tick, so the cpp-httplib check runs every
iteration and their slots are freed quickly: src/http/handlers/sse_transport_provider.cpp:184
and src/http/handlers/trigger_handlers.cpp:560. The fault stream is the exception.

There is no reaper and no other user of the tracker. SSEClientTracker
(include/ros2_medkit_gateway/core/http/sse_client_tracker.hpp:30-48) is a plain counter with no
per-connection identity, timestamp or expiry, and no timer or periodic sweep touches it.

Two smaller points in the same area:

• src/ros2_medkit_gateway/README.md:1239 says the sse.max_clients default is 10. The code
  (src/gateway_node.cpp:112), the shipped YAML (config/gateway_params.yaml:215) and
  docs/config/server.rst:641 all say 2.
• No test covers disconnect and reconnect, or the release delay. The existing slot tests
  (test/test_sse_fault_handler.cpp:284, :836) call the releaser directly, which skips the
  timing completely.