Bootstrap registration: HA client, planner, controller wiring

[as51340]

Implements specs/operator-mvp/issues/03-bootstrap-registration.md: a freshly provisioned MemgraphCluster becomes an actual HA cluster with no human action — all coordinators added, all data instances registered, one MAIN promoted.

What's in the slice

• Memgraph HA client (internal/memgraph) — a narrow interface (ShowInstances, AddCoordinator, RegisterInstance, SetInstanceToMain) plus a Connector seam, implemented over the neo4j Bolt driver. The driver is referenced nowhere else; queries mirror the HA chart's registration grammar, run as autocommit (coordinator queries cannot run in transactions), and SHOW INSTANCES rows are parsed by column name so added/reordered columns don't break parsing.
• Registration planner (internal/planner) — pure diff: declared topology + observed SHOW INSTANCES in, ordered commands out (coordinators, then data instances, then the initial MAIN promotion; empty when converged). MAIN is planned only when no MAIN exists — an existing MAIN, on any instance, is never overridden. Undeclared instances are left untouched (no unregistration in v1).
• Topology derivation (internal/resources.DeclaredTopology) — per-pod registration identity from ordinals: coordinator ordinal N is Raft coordinator N+1, data ordinal N registers as instance_{N+1} (matching the chart's naming), addresses are the pods' stable headless-Service FQDNs. The FQDN derivation is shared with the coordinator start script so advertised and registered addresses cannot drift.
• Controller wiring — after server-side-applying the workloads, reconcile gates on both StatefulSets reporting every pod ready, then finds the coordinator leader (a coordinator reporting itself leader is used directly, a follower redirects to the leader it reports, and a fresh leaderless cluster bootstraps against the first reachable coordinator, mirroring the chart), plans, and executes. Unreachable coordinators requeue on a delay rather than erroring — Bolt lagging pod readiness is a normal startup phase. After issuing commands the reconcile requeues to verify convergence rather than assuming success.

Everything is read-before-write and idempotent, so killing the operator mid-bootstrap just resumes with only the missing registrations.

Acceptance criteria

• A fresh MemgraphCluster converges to fully registered: all coordinators added, all data instances registered, one MAIN promoted
• MAIN promotion is issued only when no MAIN exists; an existing MAIN is never overridden
• Killing and restarting the operator mid-bootstrap still converges with no duplicate or failed registrations
• Planner covered by pure unit tests: fresh-cluster bootstrap, partially registered cluster, fully converged no-op, MAIN already present
• Controller registration flow covered by envtest with the HA client mocked behind its interface
• The HA client interface is the only place the Bolt driver is referenced

Testing

• Planner: pure table-driven diff cases (fresh bootstrap, partial, converged no-op, MAIN present, leaderless-but-registered, undeclared instances ignored).
• Topology: golden tests for names/IDs/FQDNs, plus a guard that the declared coordinator addresses match what the start script advertises.
• Controller: envtest with an in-memory fake cluster behind the Connector seam. The fake rejects duplicate registrations and second MAIN promotions, so any non-read-before-write behavior fails loudly. Covers: no Bolt contact before pods are ready, full fresh bootstrap with a convergence-confirming follow-up, resuming a partial bootstrap, and executing against the leader a follower reports.
• make lint, make lint-config, make test-unit, make test all green.

Live-cluster verification lands with the KinD e2e harness (issue 04).