feat(generator): discover DTOs carried as [Factory] entity properties (TRIM-002)

[keithdv]

Summary

Third plan of the TRIM todo — the second of the two zTreatment failure classes. DTO discovery correctly refused to treat [Factory] entities as DTOs, but never walked their properties, so a DTO reachable only as an entity property (TreatmentBanner on the [Execute]-opened aggregate, DashboardContactResult in a List<T> on a query entity) got no preservation and was trimmed on the client.

Changes

• Entity property-graph discovery: every class carrying [Factory] directly walks its own public property graph (inherited chain included) during generation and emits Register<T>/PreserveType<T> for reachable DTOs in its own registrar. The entity itself is never bucketed (DI registration preserves it); factory-typed properties are skipped — each [Factory] class's own registrar owns its graph, so no global reachability set and free cross-entity cycle safety.
• Namespace hardening (TRIM-001 code-review callout): IsDtoStructureCandidate's System exclusion is now a segment match — consumer namespaces like Systems.Domain are no longer silently excluded, while System/System.* framework types remain excluded (both halves pinned by tests).
• Tests: 13 new EntityPropertyDtoDiscoveryTests with per-tree registrar assertions (child-vs-parent separation), base-class inheritance, [Factory] record aggregates, LazyLoad<T> Value-descent, cycles, cross-walk dedupe. Harness: EntityPropertyDtoSmokeTest — carried DTO + record never constructed in harness code, negative control verified (walk disabled → ctor stripped → exit 1).
• Docs: docs/trimming.md + CLAUDE-DESIGN.md now document the entity-graph entry point, the manual-preservation boundary, and the deliberate interface-factory-impl exception.
• Skips the second flaky RelayTimingTests member (identical timeout signature; per user decision).

Reviews (in docs/todos/TRIM-dto-trimming-preservation-gaps/reviews/)

• Plan review: APPROVED, 5 callouts folded in — incl. correcting the draft's LazyLoad suspicion (002-plan-review.md)
• Test gate: CLEARED — should-covers closed with tests (002-test-review.md)
• Code review: no veto findings (002-code-review.md)

Test plan

• Units 581+581 → 585+585 green; integration 561+561 green (sequential run authoritative — the pre-existing FactoryEventRelay family is parallel-load flaky, all members green isolated/sequential)
• Trimmed harness green locally (win-x64), negative control verified
• This PR's CI run exercises the trimming gate on linux-x64

🤖 Generated with Claude Code