INTR-652 - Reduce startup latency cpu time and memory use in sidecars

[FarinM]

After #216 narrowed the probe and workflow children, the parent sidecar still loaded framework code it never executes. Imports for small helpers pulled in Hub database services, broad type/schema barrels, and the agent execution runtime. The supervisor also imported the aggregate control channel, which constructs every control-message validator and the combined union at import time. ArkType compiles validators as those schemas are constructed, so each sidecar paid initialization and memory costs for unused dependencies before connecting to the Hub.

This narrows the sidecar's startup graph and removes a redundant event validation pass:

• Use @intx/hub-sessions/substrate for repository access. Move workflowSourceAssetMountPath into @intx/types/workflow-sources and replace parseAgentId with the existing deriveWorkflowRunId. Deployment code calls the shared helpers directly; the injection fields and duplicate test and benchmark helpers are removed.
• Import byte, ID, credential, grant, and adapter configuration helpers through focused entry points. Tool factory checks use @intx/agent/authoring; workflow inspection uses the definition/projection entries. inlineBodyRef lives in a leaf module so the probe does not load body rewriting code through the projection entry.
• Import the individual IPC sender, receiver, and payload modules from the supervisor, recycling code, and pack client. The supervisor uses the existing per-kind validation path without constructing the combined control-payload union. Shared payload schemas and the message header remain eager.
• Validate inference events once at the IPC receiver with parseInferenceEvent and forward the canonical InferenceEvent result directly. EventPayload remains a type alias; its redundant validator export is removed.
• Extend the boot-graph tests to cover the sidecar. Positive assertions check that the walk reaches deployment wiring and the repository substrate; exclusions check for Hub services, database dependencies, unused runtime/schema barrels, and the aggregate control channel. The child and probe checks remain in the renamed process-boot-graph.test.ts.

Measurement	Before	After
Startup to WebSocket hello	724 ms	497 ms
CPU time after hello	846 ms	569 ms
RSS after hello	410.0 MiB	243.5 MiB

Measurements are medians of 12 alternating runs in fresh processes per variant on macOS arm64 with Bun 1.4.3, comparing the branch's base (f9c940190) with c9aaca55f. Each run uses an empty data directory and a local WebSocket stub. Startup is measured from process spawn to the first hello; CPU time and RSS are sampled just afterward. These cover startup with no deployments, not workflow throughput or memory under sustained load, and do not quantify the saving from removing the second event validation pass.

Removed exports have no remaining in-repository callers. Run ID validation still uses parseRunAddress; invalid-address diagnostics now use deriveWorkflowRunId's JSON-escaped formatting.

Validation: make all passed with 10,137 tests passing and two live Google API tests skipped. The first commit also passed the focused 301-test suite independently, including its guard against database imports in the sidecar.

[linear]

INTR-652

[alexanderguy]

LGTM