fix(runtime): preserve Agent Graph results across session revisions

[chinawch007]

Summary

Fixes #5331

Editing and resending a follow-up after Agent Graph completion could fail with operation_unavailable: revision preflight recognized subagent and agent_swarm results, but missed the child references in persisted JSON agent_output responses. The revision now retains those references so the edited question can run in a new version while the original version remains available.

• Recognize the defined agent_output envelope across result and diagnostic views, including historical JSON in messages, RuntimeEvents, and legacy archive bodies. Validate child ownership, retained parent turns, Graph identity, terminal Run/event identity, lineage, and Artifact ownership. An earlier running poll is accepted only when retained history also contains a terminal result for the same child, Run, and Turn.
• Preserve validated child identities within the revision family. Independent Side Conversations instead receive static result snapshots and copied Artifacts. Rewrite attachment links on subsequent copies, and migrate affected legacy archive transitions to ledger-backed snapshots with matching digests and usable pagination.
• Add regressions for JSON reference collection, invalid references, duplicate reads, polls, diagnostic views, and repeated snapshot/archive copies. Extend the production Host/provider-wire test through Graph completion, an ordinary follow-up, revision creation, and a successful edited turn; assert that child results survive, the child is not rerun, and the original follow-up remains available.

Verification

• Passed npm run lint, npm run format:check, npm run build, and npm run typecheck.

• Passed npx --no-install knip --workspace apps/desktop and npx --no-install knip --workspace packages/ui.

• Passed 72 tests across conversation-copy-agent-output, conversation-copy, session-revision-graph-references, session-revision-diagnostics, and session-revision-protocol after rebuilding.

• Passed the targeted production Host integration test:

  ✔ production Host executes and durably supervises an Agent Graph over a real provider wire
  ℹ tests 1
  ℹ pass 1
  ℹ fail 0
  

• A baseline comparison using the same historical JSON fixture collected 0 child references at 4dc6a47f9 and 1 with this fix.

• Not run: full repository npm test and manual Desktop UI reproduction. The integration test uses a local provider fixture, not a live external model.

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: OpenAI Codex — code review, regression verification, commit preparation, and this PR description. The commit includes Generated-by: OpenAI Codex.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No