feat(generator): per-assembly event-preservation registrar — subscribe-only events survive trimming (TRIM-003+007)

[keithdv]

Summary

Two plans of the TRIM todo in one PR: TRIM-003's verification (deliberately red until this fix) and TRIM-007's fix that turns it green.

TRIM-003 (verification): a FactoryEventBase-derived record whose only client-side reference is a generic Subscribe<TEvent> call site — the real consumer shape — loses its constructor under PublishTrimmed=true. The type resolves (runtime [FactoryEvent] scan works) but deserialization throws the DeserializeNoConstructor-shaped failure: inherited [DynamicallyAccessedMembers] does not flow to derived types under ILLink. The v1.4.0 "annotation covers every descendant" story was wrong, and the old smoke test couldn't see it (it constructs its event directly). Evidence triplet: red trimmed / green untrimmed / green with call-site DAM.

TRIM-007 (fix, per user decision — generator emission): a fourth incremental pipeline branch discovers every concrete, accessible FactoryEventBase descendant declared in a compilation (records-only CreateSyntaxProvider — descendants carry no attribute), walks it with the shared bucketed WalkDtoGraph (root + nested property graph), and emits one per-assembly NeatooEventPreservationRegistrar riding the existing assembly-level [NeatooFactoryRegistrar] discovery (whose DAM-annotated Type parameter roots the registrar under full trimming). Declaring an event is now sufficient — no handler, no subscription, no consumer annotation, no LinkerConfig entry. Nested reference types in event records stop being a documented manual case.

Key design points from review: accessibility gate (private/protected/file-scoped nested event records are skipped — the repo's own private test events would otherwise break every consumer build), FQN base matching, ordinally-sorted deterministic output, no emission when no events.

Verification

• TRIM-003's subscribe-only harness check green in the pure consumer shape incl. a nested record; negative control (pipeline disabled → identical failure, exit 1)
• 10 new EventPreservationDiscoveryTests (buckets, nested walk, abstract intermediates, accessibility gate, generics, FQN decoy, dedupe, determinism, no-events)
• Docs corrected end to end — trimming.md, factory-events.md, CLAUDE-DESIGN.md (incl. the now-inverted v1.4.0 supersession narrative, with history note), FactoryEventBase.cs, Design pattern files, skill references, IL2026 justification
• 595+595 unit, 561+561 integration, 0 failed; trimmed harness all checks passed

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

• 007-plan-review.md: CONCERNS → veto (accessibility gate) folded before implementation
• 007-test-review.md: gate CLEARED
• 007-code-review.md: 3 veto doc findings (surviving DAM overpromises outside the anchor list) — all fixed in-branch

Test plan

• Full solution build/test green both TFMs (build doubles as the accessibility-gate proof)
• Trimmed harness green locally, negative control verified
• This PR's CI run exercises the trimming gate on linux-x64

🤖 Generated with Claude Code