Add concurrency failure-harness checks

[lucarlig]

Summary

Adds dedicated Loom/Miri checks for concurrency and interpreter-friendly regression coverage.

This PR intentionally does not fix the issues the suite exposes. It only adds the test harness and CI jobs so the failures are visible and reproducible.

No test-only public API or internal-test-hooks feature is exposed. Loom uses a normal loom Cargo feature. Miri uses cfg(miri).

Test layout

crates/contextforge-gateway-rs-lib/src/gateway/
  session_manager.rs       # Loom concurrency model tests live beside the private helpers
  miri_checks/
    namespace_routing.rs
    config_serialization.rs

Reproduce from main

git checkout main
git pull
git checkout -b reproduce-concurrency-ci-failure
git cherry-pick 4e65148

Run the Loom check:

cargo test --locked -p contextforge-gateway-rs-lib \
  --features loom \
  gateway::session_manager::concurrency

Expected result: 3 pass, 1 fail.

Run the Miri checks:

cargo +nightly miri setup
cargo +nightly miri test --locked -p contextforge-gateway-rs-lib miri_checks

Expected result: 4 pass, 2 fail.

Regression coverage map

Every issue called out below has a regression test in this PR.

Issue	Regression test(s)	Command
Concurrent session return can erase a live backend service	gateway::session_manager::concurrency::concurrent_borrowers_do_not_erase_returned_transport	cargo test --locked -p contextforge-gateway-rs-lib --features loom gateway::session_manager::concurrency
Overlapping backend prefixes can route tool calls to the wrong backend	gateway::mcp_gateway::miri_checks::longest_backend_prefix_wins_for_tools	cargo +nightly miri test --locked -p contextforge-gateway-rs-lib miri_checks
Overlapping backend prefixes can route resource reads to the wrong backend	gateway::mcp_gateway::miri_checks::longest_backend_prefix_wins_for_resources	cargo +nightly miri test --locked -p contextforge-gateway-rs-lib miri_checks

Additional non-failing coverage in the suite:

• gateway::session_manager::concurrency::cleanup_does_not_resurrect_returned_transport
• gateway::session_manager::concurrency::different_sessions_do_not_interfere
• gateway::session_manager::concurrency::multiple_backends_preserve_unborrowed_services
• gateway::mcp_gateway::miri_checks::missing_separator_does_not_match
• gateway::mcp_gateway::miri_checks::hyphen_and_underscore_backend_names_are_distinct
• gateway::session_store::miri_checks::session_mapping_msgpack_round_trip
• gateway::session_store::miri_checks::user_session_key_msgpack_round_trip

Issues exposed

1. Concurrent session return can erase a live backend service

Concurrent calls for the same downstream MCP session can erase a live backend service in SessionManager.

Bad interleaving:

1. Request A borrows backend backend-a, so the map temporarily stores None.
2. Request B borrows while A owns the service and gets a holder containing None.
3. Request A returns the real service.
4. Request B returns None and overwrites the real service.

Observed Loom failure:

left: Some(None)
right: Some(Some(7))

2. Overlapping backend prefixes can route to the wrong backend

Backend tool/resource names are split by matching the first backend-name prefix. If one backend name prefixes another, the shorter name can be selected first.

Example:

backends: counter, counter-one
tool: counter-one-increment

Bad parse:

backend: counter
tool: one-increment

Correct parse:

backend: counter-one
tool: increment

Observed Miri test failures:

left: Some(("counter", "one-increment"))
right: Some(("counter-one", "increment"))

left: Some(("counter", "one-memo://insights"))
right: Some(("counter-one", "memo://insights"))

[lucarlig]

@dawid-nowak what do you think about miri and loom/shuttle will this tool help us or we can close and ignore?

[dawid-nowak]

I haven't looked into it to be honest. Let's leave it hanging it might be useful.

[lucarlig]

this is legacy at this point not relevant for sessionless