Add the genuinely-missing Mode A partitioning assertions for plan §14.5:
survivor stability on both 3->2 (scale-down) and 3->4 (scale-up) peer-set
resizes. The existing pod-reassignment test only checks that a removed pod's
shards are re-homed -- it does NOT pin that surviving pods keep the shards they
already owned, so a churn-inducing partitioner would pass it silently. This is
the regression parent bf-30m2j is actually worried about (two pods churning
ownership in a multi-replica HA deployment).
Enabling infra (required so the test compiles under `cargo test -p miroir-core
mode_a` with no --features):
- lib.rs: un-gate `mode_a_coordinator` and `peer_discovery` -- only the DNS
`refresh()` path needs the optional `peer-discovery` feature (gated inside
the function); the types and the test-only `set_peer_set_for_test` do not.
- mode_a_coordinator.rs: widen `set_peer_set_for_test` to
`cfg(any(test, feature = "test-helpers"))` so integration tests in other
crates can also reach it. No new feature flag -- `test-helpers` already
exists and is already used by miroir-proxy.
Tests (anti_entropy.rs, new `mode_a_minimal_reshuffling_tests` module):
- scale_down_survivor_stability: on 3->2, pod-1/pod-2 keep every shard they
owned at 3 pods; only pod-3's departed shards may change hands.
- scale_up_survivor_stability: on 3->4, exactly-one-owner + full coverage at 4
pods, and each original pod keeps its shards except those the new pod-4
takes over (no shard moves between original pods).
Verified the assertions bite: a round-robin partitioner `peers[s % N]`
satisfies exactly-one-owner but violates survivor stability (21 violations on
3->2, 30 on 3->4), so these tests fail under round-robin and pass only under
rendezvous. The in-source comment cites concrete shards (4 on 3->2, 5 on 3->4).
Full lib suite in default config: 744 passed, 0 failed.
Co-Authored-By: Claude <noreply@anthropic.com>
0ade87a3de