test(session-net): replace yield_now spin-waits with watch

[VerifiedOrganic]

Refs #578

What

Replaces the tokio::task::yield_now() spin-waits in opc-session-net that the issue identifies as dangerous, with level-triggered tokio::sync::watch signalling.

Why

A yield_now spin keeps the waiter runnable for the whole wait. Under #[tokio::test(start_paused = true)] the current-thread runtime then never goes idle, so tokio's virtual clock never auto-advances and the enclosing timeout — measured on that same virtual clock — never fires. The failure mode is a hang until the CI job's wall-clock kill, not a flake. A timeout that cannot fire is not a timeout.

Changes

• configured_ceiling_also_reserves_post_connect_rpc_time (start_paused): the accounting-settlement spin now parks on a watch generation that ConnectionAttemptTestAccounting bumps on every recorded transition (subscribe_changed).
• wait_for_cached_lane_to_empty helper (reached from the start_paused cached_consensus_reaper_never_races_an_in_flight_lane): now parks on an emptied latch the cached-connection reaper publishes when it retires the lane. The async connection mutex cannot be polled from wait_for's synchronous predicate, so the reaper publishes the edge directly.
• historical_cas_is_rejected_after_server_restart_without_redispatch: the backend poll now parks on a cas_effects watch the ReplicationDispatchSpy bumps on each committed CompareAndSetResult::Success, mirroring the file's existing CancellableStallBackend::wait_for_active convention from test(session-net): mtls_backend_deadlines flakes under workspace-level runtime oversubscription #566.

watch::Receiver::wait_for evaluates its predicate against the current value before parking, so an already-reached state resolves immediately with no lost wakeup, and a parked waiter lets the paused clock advance. Per the issue's guidance, this does not raise any deadline and does not weaken any assertion: timeout bounds (1s/1s/2s) and all post-wait assertions are unchanged.

Verification

• cargo fmt --all --check ✓
• git diff --check ✓
• cargo clippy -p opc-session-net --all-targets --all-features -- -D warnings ✓
• cargo test -p opc-session-net --all-features — 291 tests, 0 failures ✓ (the two paused tests now finish in 0.00s virtual time instead of hanging)
• Adversarial review swarm (spec/acceptance, adversarial-runtime, boundary/security): all three SHIP, no BLOCKER/MAJOR findings. The runtime reviewer independently traced the lifecycle math and confirmed every caller reaches the reaper's take(), level-triggered wakeup-safety on all three waits, and a negligible production footprint (a subscriber-less send_replace on the rare retirement path).

Scope note (priority 3)

The issue's priority 1 (the two paused-clock hangs) and priority 2 (the named three_node_quorum test) are fully addressed. Priority 3 ("the remaining instances, as cleanup, only where a test actually contends") is deliberately split out, matching how #566's scope note asked that the broad sweep be a separate decision.

A workspace-wide parse found 73 genuine tokio::task::yield_now spin-loops (68 timeout-wrapped); only the two fixed here are PAUSED-REACHABLE (confirmed hangs). The remaining ~70 span eight other crates (opc-ipsec-lb 17, opc-session-store 15, opc-session-net 11 more, opc-ipsec-xfrm 11, opc-persist 5, opc-diameter-transport 4, opc-runtime 3, opc-session-cache 3, opc-tls 1) and largely need invasive signalling — global METRICS watches crossing into opc-redaction, semaphore-permit waits, or the production opc-ipsec-xfrm/src/observation/linux.rs drain loops (537, 661) — disproportionate to a test-reliability fix. Recommend per-crate follow-up issues for that sweep.

[VerifiedOrganic]

Resolved the main merge while preserving the newer consensus regressions, then corrected the test wait in b2273ae35d3beb69bdd04dee953e153f75faa66b. Lane retirement notifications are test-only and cover both lifecycle and idle retirement. The waiter subscribes before checking the current cached connection and rechecks that state after each notification, so an earlier retirement cannot satisfy a wait for a reused lane.

Local verification passed: 89 consensus/lifecycle unit tests with four test threads, the exact historical_cas_is_rejected_after_server_restart_without_redispatch integration regression, affected all-target/all-feature Clippy, formatting, and diff checks. Complete current-head hosted validation remains required before merge.

[VerifiedOrganic]

Deferring refreshed head b2273ae35d3beb69bdd04dee953e153f75faa66b on a current-head conformance failure. The merge conflict and test-only retirement notification were repaired, preserving newer main tests and rechecking actual connection state after notifications. All 89 focused consensus/lifecycle tests, the historical-CAS restart regression, and affected Clippy pass locally.

The kernel-independent conformance job fails persistent_three_voter_consumer_write_does_not_spend_budget_on_a_read_quorum. The healthy write returns outcome_unknown within the existing operation-budget test, with delayed ReadBarrier=0 and delayed empty AppendEntries=0. This failed conformance requirement remains a merge blocker despite the passing focused wait tests.

Next step: the PR maintainer and session-net/session-store maintainers should reproduce this exact conformance failure against the refreshed branch and main, establish its cause, and land a reviewed correction without increasing the operation budget. Coordinate with #741 only if matching evidence establishes overlap. Then run every required local and current-head hosted gate. Keeping this PR open as a draft; the test failure is preserved and no passing status is inferred from a different head.