Minor findings from the stability review

[lukashes]

Small defects collected during the pre-GA review; none affects data safety, batching them to avoid issue noise.

• validator.zig prints user-facing status via std.debug.print instead of the stdout path used by main.zig (src/source/postgres/validator.zig:2,56,91,107,126).
• Stale comment in source.zig:231: says "batch wait time (~6 sec)", actual BATCH_WAIT_MS is 100ms.
• main.zig has no errdefer for source/producer between their creation and the handoff to Processor (src/main.zig:104-113) — on a wiring error the Postgres connection is not closed (the process exits anyway, so cosmetic).
• getPartitionKey swallows all errors including OOM via else |_| {} (src/processor/processor.zig:185).
• producer.zig getOrCreateTopic: topic_conf leaks when rd_kafka_topic_new fails.
• Observability.noop() leaves allocator undefined while respondMetrics reads it unconditionally (observability.zig, http.zig); unreachable with the current wiring since serve only starts when observability is configured, but fragile.
• relation_registry.zig register: column name strings leak when the final relations.put hits OOM (OOM-only path).
• Identifiers are interpolated into SQL strings in validator.zig (checkTableExists) and replication_protocol.zig (publication/slot checks). Injection is blocked by config validation upstream, but the guard lives far from the query; escape at the call site as hardening.