Share one controlled-peer test harness

Problem

Four copies of the same test harness now exist, each written for one crate's needs: silent_peer in crates/peryx-ha-distributed/tests/unit/raft/network_tests.rs, answer_beat in the same crate's beacon_tests.rs, ControlledServer in crates/peryx-upstream/tests/unit/client/deadline_tests.rs, and the bounded variant added by PR #2117.

They all do the same three things: stand up a raw peer on a real socket, run the tokio clock across the I/O rather than holding a paused clock over it, and bound the accept and read so a dial that never arrives fails instead of hanging.

The duplication is not merely repetition. Each copy resumes the clock at a different call depth, which is why the _paused-clock-contract check added in PR #2122 has to resolve calls transitively through a file to decide whether a test is safe. A single shared harness would let that check test for one known call instead of walking the graph.

Example

Three of the four deadline_tests.rs cases clear the contract check via ControlledServer::requested calling run_clock, two calls deep. The raft/network_tests.rs cases clear via silent_peer at a different depth. A fifth copy at a new depth is the failure mode this issue prevents.

Required change

Extract one harness into peryx-test-support and move the four copies onto it.

The architectural constraint is the one that has broken this crate before: a shared helper in peryx-test-support must not name a type from the crate it helps. Doing that once left peryx-storage's own tests failing with 21 type and privacy errors. Keep the shared harness to third-party types — tokio::net::TcpListener, std::time::Duration, tokio::task::JoinHandle — and let each crate compose its protocol-specific behavior on top.

The copies differ in what the peer does after accepting: one stays silent, one answers a heartbeat, one feeds a response chunk by chunk. That difference belongs to the caller. What is shared is the bind, the accept with a bound, the proof that the request arrived, and the clock discipline.

Simplify _paused-clock-contract to recognise the shared harness once the copies are gone, rather than resolving calls transitively.

Acceptance criteria

• One harness in peryx-test-support, named after what it does rather than after a crate.
• All four call sites use it, and no copy remains.
• The harness names no type from any crate that consumes it.
• Every test that used a copy still fails under the mutations its own pull request established — halved timing, doubled timing, behavior removed. A refactor that quietly weakens an assertion is the risk here.
• _paused-clock-contract recognises the shared harness directly.

Boundary

Test harness consolidation. No production change, and no change to what any test asserts. If moving a copy reveals that one of them was asserting something weaker than its siblings, say so rather than silently levelling them.

Start by reading the four harness implementations in crates/peryx-ha-distributed/tests/unit/raft/network_tests.rs, its beacon_tests.rs, crates/peryx-upstream/tests/unit/client/deadline_tests.rs, and the bounded variant from PR #2117, then inspect peryx-test-support and _paused-clock-contract. Move only the shared socket, bounded accept/read, arrival proof, and clock-discipline behavior; preserve each protocol-specific callback and its existing mutation-tested assertions. Done means all four sites use one dependency-neutral harness and the contract check recognises it directly.