feat(desktop): show Agent Graph output previews

[testikun]

Summary

• Project a 280-code-point per-operator output preview and usage-derived token rate from the child Session's existing RuntimeEvent/read-model authority. Live deltas fold replay-safe absolute offsets; durable rebuilds use completed model text and provider-reported output-token usage.
• Carry the optional bounded fields through the strict Runtime Host allowlist/decoder, bumping the compatibility epoch from 109 to 110, and render distinct live/result labels plus TPS in the Desktop Agent Graph panel.
• Keep the existing Open child task path and leave graph scheduling, admission, execution, recovery, and child-session semantics unchanged.

Fixes #3714

Verification

• node --test packages/runtime/dist/__tests__/stream-graph-projection.test.js packages/runtime/dist/__tests__/stream-graph-read-model.test.js packages/runtime-host/dist/__tests__/agent-graph-protocol.test.js packages/runtime-host/dist/__tests__/agent-graph-coordinator.test.js apps/desktop/dist/main/__tests__/agent-graph-panel.test.js — 45 passed
• node --test --test-name-pattern="boots an empty graph" packages/runtime/dist/__tests__/stream-graph-coordinator.test.js — the CI regression passes with bounded output-commit and exact invalidation coverage
• npm --workspace @maka/runtime run test:dist — 3,224 tests; 3,211 passed, 13 skipped, 0 failed
• npm run typecheck — passed; affected Runtime, Runtime Host, and Desktop typechecks also passed after rebasing onto current main
• npm run lint — passed
• npm run format:check — passed
• npm --workspace @maka/desktop run build:with-deps — passed after the final rebase
• npm --workspace @maka/desktop run check:architecture — 71 passed; renderer architecture check passed against the final base
• npx knip --workspace apps/desktop and npx knip --workspace packages/ui — passed
• npm --workspace @maka/desktop run build-storybook — passed
• node scripts/protocol-epoch-check.mjs --staged — passed (109 -> 110)

Real-window verification: built and launched current head ff866d713 in a visible local Electron App, opened the production Agent Graph panel through the repository's persisted Host projection, and verified streaming/completed output previews, measured TPS, the 24-operator scroll layout, collapse behavior, and retained Open child task actions.

Real App screenshot

This is a direct capture from the locally built and running Maka Electron App after the test above; it is not a Storybook render, mockup, or generated image.

Review focus

RuntimeEvents remain the only output/usage authority. The new projection is presentation-only and bounded before it reaches the wire; it does not create graph activity records for text deltas or participate in supervisor replay.

AI use

Select exactly one:

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

Tool(s) and scope: OpenAI Codex implemented the bounded Runtime/read-model projection, strict protocol plumbing, Desktop rendering, deterministic tests, Storybook fixture, and verification. The material commit carries the required Generated-by: OpenAI Codex trailer.

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