fix(images): refresh trixie package pins - #11634

…11633)

## Outcome

OpenClaw and Hermes can build their normal images and start providerless
sandboxes without a selected model or inference route. Configuration
represents absent inference explicitly until the existing managed
inference lifecycle supplies a real route and model.

## Reason

Providerless onboarding passed an empty model through Dockerfile
preparation. OpenClaw rejected the generated empty model ID; Hermes
rejected the missing `NEMOCLAW_MODEL`. The managed startup profile also
substituted a placeholder model, which did not represent a configured
inference route.

This follows #11584. Both generator failures were reproduced against
main revision `c356332eae2c5bcc03fd64933db704b91795d980` before the
repair.

### Related issues

Refs #11548.

## Changes

- Preserve explicit inference absence through Dockerfile arguments, both
real configuration generators, and the internal startup profile.
OpenClaw and Hermes accept an absent profile; other agents retain their
existing requirements.
- Share the absence check between generation and startup so partial
inference input still fails ordinary validation. Real-generator
regression tests cover absent, configured, partial, and deferred
configuration.
- Retain normal image entrypoints, plugin integration, Hermes managed
restrictions, and dashboard seeding. Handle nullable startup profiles in
clone, rebuild, and export consumers.
- Exercise the existing managed inference configuration updater after
providerless generation. Failed route changes preserve the absent
configuration; invalid Hermes dashboard routing does not replace
existing state.

The E2E fixture validates restart inputs before stopping the gateway. It
restarts a selected user service, or calls the candidate CLI's existing
startup code for the registered gateway. A failed selected-service
restart retains that service for cleanup. Gateway health, sandbox
readiness, retained state, and agent turns remain separate assertions.
Unsupported OpenShell start/stop commands are removed from this fixture.
Production recovery behavior, network preparation, component schemas,
authentication, policy ownership, activation deadlines, and retries
remain unchanged.

## Verification

- Gateway fixture and receipt repair
`41c99ba87bf3e2f4c2e8e11525cd9324b6950d9f`: 88 E2E support tests, 18
receipt-refresh tests, and 15 candidate-authority tests passed locally
and on Linux from this committed source. The subprocess tests execute
the restart script with controlled registry/startup modules. Live
gateway startup also passed in the CI activation run below. CLI
type-checking and all repository checks passed, including the Pi receipt
gate.

- `npx vitest run --project integration
test/generation/providerless-agent-config.test.ts` — 33 passed on Linux,
zero skipped. The corrected test helper in
`0427b021a033dc79999386346a6928ca18085e73` supplies the selected
provider and agent identity to the real Dockerfile patcher. Its bytes
match the Linux test run.
- Affected startup, Dockerfile, clone, rebuild, and build-context suites
— 1,023 passed across 41 files. Agent generator, Hermes
policy/dashboard, and affected E2E support suites — 378 passed across
eight files. These runs overlap with other recorded runs; counts are not
additive.
- Full managed inference action suite — 295 passed across 20 files.
External route/service operations in these command tests are mocked.
- `NODE_OPTIONS=--max-old-space-size=8192 npm run typecheck:cli` —
passed.
- `npm run build:cli` and `npm --prefix nemoclaw run build` — passed.
- Linux AMD64 live validation of production commit
`903aa4a2045b8494df6f286ed17361c2c8e3b920` with OpenShell 0.0.116 — both
normal image builds passed; both sandbox supervisors reached `Ready`;
both agents started and passed health checks. The gateway had zero
providers before and after testing.
- Live OpenClaw configuration had no providers and no primary model; the
NemoClaw plugin remained enabled. Hermes had no routing keys and
retained manual approvals. OpenClaw's observed pre-install plugin
notices were nonblocking.
- Exact image IDs: OpenClaw
`sha256:4353b469eacb7f8d7cc04e6481c715077c51f9574ae2ab65874493a73a3809ec`;
Hermes
`sha256:ecd37be5e607857d96fd79e010b4d8712582cc8cef2c572266aeeb1ebcebad56`.
Runtime versions were OpenClaw 2026.7.1 and Hermes 0.20.6. Both retained
user `sandbox` and entrypoint `/usr/local/bin/nemoclaw-start`.
- Test sandboxes, their isolated gateway, and their empty test network
were removed after evidence collection. Unrelated workloads and retained
failure evidence were preserved.
- Secret scanning passed; the diff contains no secrets, API keys, or
credentials.

The Pi receipt gate passes locally. Both native Pi jobs in
[managed-image run
34727786488](https://github.com/NVIDIA/NemoClaw/actions/runs/34727786488)
built, published, and validated their images successfully. The
checked-in receipts preserve the downloaded contract bytes for source
`e44d6f1ee65e05bfab0949f01a91be31f497aaaf`, cohort
`ghrun-34727786488-1`; the authority contains their SHA-256 digests. Pi
image inputs are unchanged between that qualified source and this
repair. Merged package dependency #11634 also cleared the Deep Agents
Code staging-base failure.

CI on `41c99ba87bf3e2f4c2e8e11525cd9324b6950d9f` completed without
failed checks. The [CI
suite](https://github.com/NVIDIA/NemoClaw/actions/runs/34730058075)
passed all CLI shards, plugin and installer tests, type-checking, and
repository checks. The [managed-image
run](https://github.com/NVIDIA/NemoClaw/actions/runs/34730057979) passed
all three direct image/startup checks, both Pi image qualifications,
staging-base validation, and both OpenClaw MCP discovery passes.

The [managed runtime activation
job](https://github.com/NVIDIA/NemoClaw/actions/runs/34730057979/job/103653850532)
passed for OpenClaw, Hermes, and Deep Agents Code using published images
from this exact source and cohort `ghrun-34730057979-1`. Its artifact
records 12 agent turns, zero Dockerfile builds during activation,
onboarding, gateway restart, reconciliation, retained-state checks, and
destruction. Cleanup completed without failures. Agent turns used the
test inference responder; this does not qualify the providerless
external-component flow or a production inference service. This result
resolves the earlier gateway restart fixture failure.

## Review notes

Sensitive paths include `scripts/generate-openclaw-config.mts`,
`agents/hermes/**`, `src/lib/onboard/**`, candidate receipt authority,
and the reviewed startup runtime bundle. Self-review and the listed
tests cover the changes. The fixture addresses the selected-service
restart and missing-input review findings. CodeRabbit reviewed changes
through `41c99ba` and reported no new actionable comments. All nine [PR
Review Advisor
specialists](https://github.com/NVIDIA/NemoClaw/actions/runs/34730704378)
completed successfully; their artifacts were collected and checked
against this commit. Advisor findings concern the deferred rebuild
lifecycle and public Hermes guidance described below. The human
change-request review remains unresolved. Earlier bootstrap commits used
the disclosed Pi receipt exception; the final receipt commit also needed
that pre-commit exception because the gate compares committed history.
The committed receipt pair now passes the gate. The PR is not ready to
merge while validation or review remains unresolved.

Review also requests rebuilding and automatic upgrades while inference
is deferred. Existing rebuild preflight refuses before deleting the
sandbox when provider/model data is absent. That broader lifecycle
change remains pending a maintainer scope decision. The nullable
prior-profile reads support rebuild after inference is configured later.

Advisor also identified public providerless guidance that is restricted
to OpenClaw. The Hermes procedure and deferred-inference explanation are
deferred to `Docs / Post-Merge Catch-Up`, as permitted for `docs/**` in
an ordinary code PR. The owning repository guidance remains in this PR.
This documentation follow-up must retain the live qualification limits
below.

The earlier providerless Linux validation covered the normal
generated-image path with capability union disabled. It did not exercise
protected managed-image root application, external-component activation,
or a successful managed inference request. Both normal policies used
their existing `best_effort` mode. Supervisors reported three skipped
filesystem rules; `/app`, `/run/nemoclaw/managed-startup-ca-bundle.pem`,
and `/run/nemoclaw/managed-startup-runtime.env` were absent. These
results do not qualify a stricter external policy. No placeholder paths
or policy changes were introduced.

The shared external-component installation could not support a second
independent lifecycle for this test. Existing declarations were not
reused or reset. Neither agent was inference-ready: a matching route,
valid model configuration, and a successful managed inference request
still need live qualification. Ordinary provider-backed behavior and
deferred configuration are covered by the listed deterministic tests.

---
Signed-off-by: Apurv Kumaria <akumaria@nvidia.com>


<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **New Features**
* OpenClaw and Hermes can now start without an inference provider or
model configured.
* Inference routing can be added or updated later through managed
configuration.
* Providerless configurations preserve supported dashboard behavior and
allow secure later route updates.
  * OpenClaw heartbeat sessions now run in isolation.

* **Bug Fixes**
* Improved handling of incomplete or missing inference settings during
onboarding, cloning, rebuilding, and dashboard configuration.
* Added validation to reject partial routing configurations and prevent
invalid credentials or model settings.

* **Documentation**
* Added guidance for providerless startup, later inference activation,
validation, and recovery behavior.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Signed-off-by: Apurv Kumaria <akumaria@nvidia.com>