[bmscomp]

Merge #441 first. Rebased onto #441, which is rebased onto main after #438, #440 and #439. Until #441 lands, the diff here includes its commit.

Problem

FaultSpec.Builder defaulted targetNamespace to the literal "kafka" (chaos/FaultSpec.java:83). The safety guard lets a step target only kates.chaos.kafka.namespace and the Kates API's own namespace (DisruptionSafetyGuard.java:140-149).

So with Kafka anywhere but kafka, every fault that names no namespace was aimed at kafka, and the guard refused it. That covers ad-hoc plans, the playbooks, the templates and the resilience scenarios. #441 makes the chart set kates.chaos.kafka.namespace from where Kafka runs, which made this the next thing to break.

Fix

• Backend. The builder's default reads kates.chaos.kafka.namespace through ConfigProvider (default kafka). A spec that names a namespace keeps it.
• Why the builder reads config itself. Specs are built by Jackson from REST bodies and stored plans, by the playbook and template catalogs, and by the resilience scenarios. None of those can inject config into the builder, and a missed path would then get a null namespace, which fabric8 may resolve to the client's own namespace. The read happens when a Builder is constructed. All of those paths do that per request, at runtime.
• MCP preview_disruption, which assumed the literal:
  • A step that leaves targetNamespace out is in the namespace the dry run reads, so it no longer draws dry-run-other-namespace.
  • outsideAgentLimits compares a step's namespace with the Kafka namespace the topology names (kafka when unknown), not with kafka. A playbook comes back as the backend built it, so with Kafka in kafka-prod its steps name kafka-prod. Compared with kafka, every playbook step would have counted as beyond the agent limits.
  • The caveat says kates.chaos.kafka.namespace. Its two FaultSpec.java refs are re-pointed and anchored, and two plain comment refs are re-pointed.
• Book. Chapter 07, Targeting Pods, says the same.

Tests

• FaultSpecTest.absentNamespaceIsTheConfiguredKafkaNamespace registers a config with namespace streaming for a fresh context class loader. It checks the builder, JSON without the field, and an explicit namespace. It fails on the old default (expected: <streaming> but was: <kafka>).
• TestMCPPreviewAgentLimitsKafkaNamespace uses a topology in kafka-prod. A playbook step in kafka-prod is within limits; one in kafka is not. It fails on the old comparison.
• TestMCPPreviewOtherNamespace now expects no caveat for a step without a namespace when the dry run names kafka-prod.

Local checks

• In kates/:
  • ./mvnw test -Dquarkus.grpc.codegen.skip=true -Dquarkus.http.test-port=0 passes on the rebased tip: 1,496 tests in 141 classes, 0 failures, 1 skipped.
  • spotless:check (JDK 25) passes.
  • ITs and the Docker and native builds run in CI only.
• In cli/: go vet ./..., go test ./... (including TestMCPCaveatRefAnchors), golangci-lint run --new-from-rev=origin/main (0 issues), govulncheck (none), check-cli-style.sh, check-cli-compat.sh.
• Book: check-book-style.sh, check-book-contrast.py, gen-version-matrix.sh --check, gen-chart-table.sh --check. The change is one sentence, so the book was not rendered.

Upgrading

A disruption schedule stores its plan as the backend parsed it (DisruptionScheduleResource.java:110,146). A schedule created before this change keeps "targetNamespace":"kafka" in the plan it stored. Recreate it to pick up another Kafka namespace.

Not in this PR

• The consumer-isolation playbook names targetNamespace: kates, the Kates API's namespace on a default install. With Kates elsewhere, kates-platform included, the guard refuses that step.