Replace the DCode Nemotron Ultra source patch with a managed profile plugin

[ericksoa]

Context

PR #6416 moved the managed DCode image to deepagents-code[nvidia]==0.1.34 / deepagents==0.7.0a6 and consumes the released Nemotron 3 Ultra harness profile. NemoClaw currently modifies the installed Deep Agents bootstrap at image-build time so that profile also resolves for the two model keys used by our managed OpenAI-compatible route:

• openai:nvidia/nemotron-3-ultra-550b-a55b
• openai:nvidia/nvidia/nemotron-3-ultra

That source mutation is unnecessary on our side. Deep Agents already discovers harness-profile plugins through its deepagents.harness_profiles entry-point group. NemoClaw can install a first-party managed plugin that aliases the released canonical profile without modifying any third-party package file.

Scope

• Add a NemoClaw-owned harness-profile entry-point plugin to the managed DCode image.
• Register both managed keys against the released canonical nvidia:nvidia/nemotron-3-ultra-550b-a55b profile.
• Keep exact DCode/Deep Agents version checks plus official native-profile and bootstrap source hashes.
• Make registration atomic, idempotent, and fail closed on version, source, canonical-profile, or partial-state drift.
• Delete patch-nemotron-ultra-profile.py, its Docker build wiring, and the patched-bootstrap digest.
• Update dependency review and focused tests to prove the official Deep Agents bootstrap remains byte-for-byte unmodified.
• Keep the existing managed model IDs and public CLI/configuration behavior unchanged.
• Preserve the existing DCode hardening, TUI first-run behavior, progressive disclosure, MCP handling, and Relay observability integration.

Acceptance criteria

• The image installs and discovers the NemoClaw-owned profile plugin through the supported harness-profile entry-point group.
• Both managed keys resolve all 12 native Ultra middleware entries.
• An unrelated OpenAI model receives no Ultra middleware.
• A complete Deep Agents graph compiles.
• Parser-repaired and native structured execute calls dispatch identical command arguments exactly once and produce the same managed shell approval/denial result.
• The image no longer copies or executes patch-nemotron-ultra-profile.py.
• The installed native profile and Deep Agents bootstrap match the reviewed official wheel byte-for-byte after the image build.
• Plugin failure-state tests cover exact-version/hash rejection, missing canonical profile, partial state, rollback, and idempotence.
• A DCode-only CI negative build strips both upstream distributions and proves the production image stops at the isolated import gate before the later dependency check.
• Dependency review, image-contract tests, focused profile tests, and the ubuntu-repo-cloud-langchain-deepagents-code E2E target pass on the final SHA.
• Direct/login-shell headless inference and interactive TUI acceptance remain green.
• No CLI, configuration schema, managed model ID, or external repository change is introduced.

Evidence

• Implementation: refactor(dcode): replace Nemotron source patch with profile plugin #6431 at 74750c5ca2830a8b92f608897900568609a0b9de
• Required DCode target: https://github.com/NVIDIA/NemoClaw/actions/runs/28919113203
• Advisor-required cloud onboarding: https://github.com/NVIDIA/NemoClaw/actions/runs/28919114738/job/85792327521

Lifecycle

Keep #6431's localized NEMOCLAW_DCODE_PROFILE_IMPORT_GATE and stripped-base regression gate until the image installation path validates deepagents/deepagents-code dependency completeness atomically. Remove the marker, workflow gate, fixture, and focused tests when that condition is met.

Out of scope

• Changing the managed model IDs.
• Vendoring or copying the native Ultra profile.
• Modifying or contributing to an external repository.
• Floating to an unreleased dependency.

Related: #6416