refactor(compute): support external driver parity

[drew]

Summary

Make external compute drivers follow the same public gRPC boundary as in-tree drivers. Canonical driver names can be supplied over UDS, driver-requested callback listeners are validated generically, and omitted process identity remains omitted so every runtime can apply its native defaults.

Related Issue

Part of #2417

Stacked on #2743.

Changes

• preserve omitted process user and group fields by default for every compute driver
• avoid adding capability messages, feature enums, or configuration for process identity behavior
• structurally validate driver-requested callback listeners independently of driver name
• allow canonical built-in names with --compute-driver-socket, with the explicit endpoint taking precedence
• route shutdown, startup, lifecycle, and policy operations through the public compute-driver RPC boundary
• add fake external UDS coverage using a canonical driver name
• document the supported extension boundary in architecture notes and agent diagnostics; no public docs/ changes are included

Testing

• mise run pre-commit
• focused shutdown/startup reconciliation tests
• process-identity create/get and policy-backfill tests
• external UDS lifecycle tests

Checklist

• Changes are scoped to the accepted issue
• Tests cover the external UDS path
• Architecture notes and agent diagnostics are updated
• Public-facing documentation changes remain in refactor(compute): unify gateway restart reconciliation #2743, not this follow-up

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-2744.docs.buildwithfern.com/openshell

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[elezar]

Features are essentially binary flags, correct. Do we see the need for something richer?