Context
From the first complete full mutation-testing run (ADR-007 shared workflow, run 29074771985, 2026-07-10, main @ 94a5d44). The client pool's shutdown and rotation logic can be mutated without any test noticing.
Surviving mutants
src/client/pool/client_pool.rs:149 — replace ClientPoolInner::is_shutdown -> bool with falsesrc/client/pool/client_pool.rs:187 — replace > with == in ClientPoolInner::ordered_slotssrc/client/pool/scheduler.rs:210 — replace PoolScheduler::notify_shutdown with ()src/client/pool/scheduler.rs:101 — replace has_waiters -> bool with false / replace || with &&src/client/send_streaming.rs:92 — replace SendStreamingConfig::chunk_size -> Option<usize> with None / Some(0) / Some(1)
Analysis
No test closes a pool and then attempts an acquire, so is_shutdown can be forced to false and both try_acquire_immediately and acquire_for_handle proceed as if the pool were live. Every pool test uses pool_size(1), where rotate_left(0) is a no-op, so round-robin slot rotation is never observed. notify_shutdown drains queued waiters with disconnected(); nothing blocks a waiter across a shutdown, so an empty body survives (such waiters would hang). The public SendStreamingConfig::chunk_size getter has zero callers (internal code reads the field directly), so all three accessor mutants survive — note issue #570's remark that the neighbouring chunk-size mutants were caught applies to the streaming path, not this getter.
Suggested tests
pool.close().await then handle.acquire().await — assert ClientError::disconnected() is returned promptly (kills is_shutdown).- Build a pool with
pool_size(2) and max_in_flight_per_socket(1); acquire and drop leases repeatedly and assert the test server observes two established connections (kills the ordered_slots rotation mutant, which leaves every acquire on slot 0).
- Saturate the single slot, spawn a blocked
handle.acquire(), close the pool, and assert the blocked acquire resolves to Err(disconnected) within a bounded timeout (kills notify_shutdown).
- Assert
SendStreamingConfig::default().with_chunk_size(4096).chunk_size() == Some(4096) and that the default is None (kills the accessor mutants).
Accepted
scheduler.rs:101 (has_waiters, both variants) is a private race-recovery guard distinguishable only under an adversarial interleaving (waiter enqueued between take_next_waiter and the losing kick CAS). Example-based tests cannot reach it deterministically; a loom model of the scheduler would be the appropriate adversary. Accepted for now.
Context
From the first complete full mutation-testing run (ADR-007 shared workflow, run 29074771985, 2026-07-10,
main@ 94a5d44). The client pool's shutdown and rotation logic can be mutated without any test noticing.Surviving mutants
src/client/pool/client_pool.rs:149—replace ClientPoolInner::is_shutdown -> bool with falsesrc/client/pool/client_pool.rs:187—replace > with ==inClientPoolInner::ordered_slotssrc/client/pool/scheduler.rs:210—replace PoolScheduler::notify_shutdown with ()src/client/pool/scheduler.rs:101—replace has_waiters -> bool with false/replace || with &&src/client/send_streaming.rs:92—replace SendStreamingConfig::chunk_size -> Option<usize> with None / Some(0) / Some(1)Analysis
No test closes a pool and then attempts an acquire, so
is_shutdowncan be forced tofalseand bothtry_acquire_immediatelyandacquire_for_handleproceed as if the pool were live. Every pool test usespool_size(1), whererotate_left(0)is a no-op, so round-robin slot rotation is never observed.notify_shutdowndrains queued waiters withdisconnected(); nothing blocks a waiter across a shutdown, so an empty body survives (such waiters would hang). The publicSendStreamingConfig::chunk_sizegetter has zero callers (internal code reads the field directly), so all three accessor mutants survive — note issue #570's remark that the neighbouring chunk-size mutants were caught applies to the streaming path, not this getter.Suggested tests
pool.close().awaitthenhandle.acquire().await— assertClientError::disconnected()is returned promptly (killsis_shutdown).pool_size(2)andmax_in_flight_per_socket(1); acquire and drop leases repeatedly and assert the test server observes two established connections (kills theordered_slotsrotation mutant, which leaves every acquire on slot 0).handle.acquire(), close the pool, and assert the blocked acquire resolves toErr(disconnected)within a bounded timeout (killsnotify_shutdown).SendStreamingConfig::default().with_chunk_size(4096).chunk_size() == Some(4096)and that the default isNone(kills the accessor mutants).Accepted
scheduler.rs:101(has_waiters, both variants) is a private race-recovery guard distinguishable only under an adversarial interleaving (waiter enqueued betweentake_next_waiterand the losingkickCAS). Example-based tests cannot reach it deterministically; a loom model of the scheduler would be the appropriate adversary. Accepted for now.