fix(performance): compile the example's tests again and gate them per PR - #900

Problem

The scheduled performance workflow has failed every night since 2026-10-13. Two causes, one masking the other:

c4a5be6 (PIP-460 M1 stream consumer) removed MessageId::segment_id. The scalable-topics block in crates/magnetar/examples/performance.rs still assigned it, so the example's test module stopped compiling:

error[E0609]: no field `segment_id` on type `magnetar::MessageId`
   --> crates/magnetar/examples/performance.rs:618:23

Reproduced locally with the exact contracts command: cargo test -p magnetar-driver --example performance --release --features moonpool,scalable-topics --locked → exit 101.

Fix

• Remove the dead block. It asserted that marker_position_matches separates ids differing only by segment; MessageId no longer carries a segment by design (see its doc comment), and the example drives no scalable workload. marker_position_matches is unchanged.
• Close the gap that let c4a5be6 merge green: no per-PR job compiled an example's #[cfg(test)] module (clippy --all-targets checks examples as binaries; the workspace test run skips examples). The non-e2e test job now runs cargo test -p magnetar-driver --all-features --locked --examples, and CONTRIBUTING.md / CLAUDE.md carry the same line in their validation chains.

Verification

• Exact contracts command, fixed tree → test result: ok. 4 passed; 0 failed.
• New CI step on the pre-fix example → exit 101, same E0609 (the guard is seen red); on the fixed tree → 4 passed in ~100 s.
• CONTRIBUTING.md $FEATURES variant → 4 passed.
• cargo +nightly fmt --all --check, cargo clippy -p magnetar-driver --all-features --examples --profile test -- -D warnings, prettier, actionlint (only the pre-existing ubuntu-26.04 label notice, identical on main).

Second masked cause (found by the first dispatch, run 41624715389)

With the example compiling, prepare reached its Python contracts and failed one test (CI python-contracts.stderr: FAILED (failures=1)):

AssertionError: 'fixture-required' != 'fixture-free' : pulsar_fixture_contract.rs

c0d62a4 (landed in the same 0cfbd30 range as c4a5be6) added a_standalone_binary_port_is_refused_before_a_proxy_starts, which reached the proxy guard through Pulsar::start_proxy(). The Docker-free binary stayed container-free only because the assertion panicked first, and the fixture policy reads that .start_proxy() call as a fixture start. The guard now lives in Pulsar::proxy_command() (the proxy counterpart of command()), start_proxy() calls it before creating a container, and the test calls it as a plain #[test].

• Classifier contract on the old test file → FAILED (failures=1) with the same assertion; on the new one → OK.
• cargo test -p magnetar-driver --all-features --locked --test pulsar_fixture_contract → 22 passed; clippy --tests -D warnings clean.

Still to prove

A second manual performance dispatch (run 41626712212, head fbde666) is running to show whether the 2026-10-13 worker failure still exists. Note the first green nightly after merge will measure from a154636 (the last successful nightly's head), so a wide comparison range is expected, not a regression.