maintenance: scope AI conversations by creator

[zqr10159]

Summary

• enforce the authenticated creator across conversation and SOP schedule create/list/get/update/delete/toggle operations
• validate the persisted schedule owner against the target conversation before execution and again before result or error delivery
• prevent protected monitor creation from loading secure-form data through an unscoped conversation lookup
• disable ownerless legacy schedules, add creator-oriented indexes, and document the manual recovery path
• keep background execution independent of a request-thread subject while persisting the validated owner on generated messages

Upgrade impact

• conversations with a null creator remain isolated from all user queries; no ownership is guessed automatically
• migration V182 disables schedules with a null or blank creator while preserving the rows for administrator review
• schedules with a missing conversation or a creator mismatch are disabled before they can execute
• the upgrade guide describes how to verify and restore an intended owner without bulk assignment

Validation

• regression proof on the previous head: four creator-boundary tests failed because the service trusted a request creator and allowed cross-creator get/list/delete
• ./mvnw -pl hertzbeat-ai -am -Dtest=SopScheduleServiceImplTest,SopScheduleExecutorTest,ConversationServiceImplTest,ChatConversationTest test -DskipITs -Dsurefire.failIfNoSpecifiedTests=false -DfailIfNoTests=false
• ./mvnw -pl hertzbeat-ai -Dtest=MonitorToolsImplTest,SopScheduleServiceImplTest,SopScheduleExecutorTest,ConversationServiceImplTest test -DskipITs -Dsurefire.failIfNoSpecifiedTests=false -DfailIfNoTests=false (20 tests)
• ./mvnw -pl hertzbeat-ai test -DskipITs (50 tests)
• ./mvnw -pl hertzbeat-startup -am -DskipTests package
• H2 fresh-migration proof: V181 followed by V182 preserved the owned schedule, disabled the ownerless schedule, and created the composite index
• repository no-hardcoded-CJK scan and git diff --cached --check

AI assistance: used for draft implementation and test iteration.
Human validation: focused ownership tests, the complete AI module suite, a fresh H2 migration proof, and the startup source package reactor all completed successfully; Checkstyle reported no violations.
Risk notes: legacy ownerless conversations and schedules require deliberate administrator recovery after verifying the target conversation owner; the migration does not guess or bulk-assign ownership.

[zqr10159]

Author remediation update:

The ownership boundary now includes conversations, messages, stream and security-data reads, plus every SOP schedule create, list, get, update, delete, and toggle path. A schedule persists its owner, creation verifies the target conversation, and background execution revalidates that persisted relationship without a login-time subject. Historical null creators use an explicit migration and isolation path with supporting indexes.

Two-user IDOR and scheduled-message injection regressions are covered. Focused AI tests (50) and the startup reactor (24 modules) passed locally; the current GitHub backend, E2E, docs, license, and label checks are green. The author-side blocker is resolved; maintainer review is still required.

[Duansg]

Scoping these by creator is the right call. On master, ConversationServiceImpl reaches conversations through findById() / findAll() with no owner check, and MonitorToolsImpl.java:312 loads a conversation's securityData — AES-encrypted monitor credentials — through that same unscoped lookup, so any authenticated user could reach another user's conversation and the secrets attached to it. The DAO-level findByIdAndCreator / findByConversationIdAndCreator approach closes that cleanly, and re-validating ownership in getScheduleForExecution before both result and error delivery is a nice touch for the background path.

Two things I checked that are not problems, so they don't get re-raised:

• The migration is type-correct. V181 creates hzb_sop_schedule with enabled SMALLINT on PostgreSQL and TINYINT on MySQL — not boolean — so SET enabled = 0 is valid on both, and creator VARCHAR(64) already exists in the original CREATE TABLE, so V182 has a column to filter on. Flyway also runs ahead of Hibernate, so the new @Index won't collide with the one V182 creates.
• creator is set explicitly even though @createdby would populate it. Both entities have AuditingEntityListener and JpaAuditorConfig reads the same SurenessContextHolder subject, so the values agree — and setting it in the builder is necessary for the ownership check inside the same transaction. Redundant but correct.

What needs a maintainer decision

The PR discloses this, so it isn't a hidden bug, but it's the part I'd want agreed before merge:

• Administrators lose access to other users' conversations entirely. Every other resource in HertzBeat (monitors, notice rules, bulletins) is globally visible, so AI conversations become the one creator-private resource type. That's a defensible choice, just an inconsistent one worth making deliberately.
• Existing conversations with creator = null become unreachable for everyone, and V182 disables their schedules. Recovery is documented as manual SQL against hzb_ai_conversation and hzb_sop_schedule.

For an operator upgrading a running instance, "your AI history is gone unless you hand-edit the database" is a rough landing. An admin-facing recovery path in the app — list ownerless conversations, assign an owner, re-enable the schedule — would turn this from a support ticket into a UI action. Happy to see it as a follow-up rather than in this PR, but it'd be good to have a plan on record.

Minor

requireCurrentUserId() is duplicated verbatim in ConversationServiceImpl and SopScheduleServiceImpl. The subject lookup already appears in several places (JpaAuditorConfig, AccountServiceImpl, AuthTokenController), so a single shared helper would be worth extracting while this is fresh.

Also, IllegalStateException("No authenticated user") will likely surface as a 500 rather than a 401 — worth checking how the controllers map it.