docs: add v0.0.96 release entry

docs/changelog/2026-07-25.mdx

@@ -0,0 +1,50 @@
+{/*
+ * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */}
+
+## v0.0.96
+
+NemoClaw v0.0.96 adds persistent baseline network policy exclusions, DNS-backed HTTPS inference switching, and host-managed default OpenShell gateways.
+It also improves onboarding and recovery diagnostics, locks and updates managed sandbox image dependencies, reduces Hermes image size, and strengthens release validation.
+
+- `policy exclude` and `policy restore` now persist an operator-approved removal of one exact agent baseline entry across rebuild and snapshot restore.
+  The commands preview the removed endpoints and affected features, reserve excluded keys, protect `managed_inference`, and require another review after agent or baseline drift.
+  `policy list`, `policy explain`, `status`, `doctor`, snapshot, and rebuild output report active or inconsistent exclusions.
+  For more information, refer to [Network Policies](/user-guide/openclaw/reference/network-policies) and the [NemoClaw CLI Commands Reference](/user-guide/openclaw/reference/commands).
+- `nemoclaw inference set --endpoint-url` now routes public DNS-backed HTTPS custom endpoints through a host-side HTTPS Pin Runtime adapter.
+  Each route receives a separate sandbox-facing credential, while the upstream endpoint and credential remain host-only.
+  OpenAI-compatible HTTP endpoints on exact loopback hosts `localhost`, `127.0.0.1`, or `[::1]` can use no authentication on ports `8000`, `11434`, or `11435`, and URL-form Node.js requests no longer forward the internal provider marker.
+  A failed compatible-provider update attempts to restore the previous provider and model.
+  For more information, refer to [Meet Custom Endpoint Security Requirements](/user-guide/openclaw/inference/custom-endpoints/custom-endpoint-security), [Set Up an OpenAI-Compatible Endpoint](/user-guide/openclaw/inference/custom-endpoints/set-up-openai-compatible-endpoint), and [Switch Inference Providers](/user-guide/openclaw/inference/manage-inference/switch-providers).
+- The default OpenShell gateway on port `8080` now uses a validated host service on Linux and Apple Silicon macOS when the platform provides the supported service manager.
+  NemoClaw-managed custom ports retain the detached-process lifecycle, while a declared external supervisor retains authority for its gateway.
+  Onboarding now uses deadline-based readiness waits and omits TLS Server Name Indication for IP-literal health probes.
+  It accepts a positively identified Docker engine even when `ProductLicense` is `Apache-2.0`, and rejects a macOS Podman compatibility socket before gateway launch.
+  For more information, refer to [Architecture Details](/user-guide/openclaw/reference/architecture), [Platform Support and Launch Claims](/user-guide/openclaw/reference/platform-support), and [Troubleshooting](/user-guide/openclaw/reference/troubleshooting).
+- Status and upgrade checks now scope gateway failure and agent-version probes to the sandbox's recorded gateway.
+  `doctor` reports incomplete lifecycle metadata without exposing stored values.
+  An explicit forced rebuild can preserve managed MCP configuration after the old sandbox loses exec access, with exact gateway, policy, and provider proofs before deletion.
+  Rebuild rechecks the canonical sandbox and recorded gateway at the delete edge, restores MCP state if that target drifts, and keeps shields unlocked when an accepted deletion remains unconfirmed.
+  Shields failures return exit code `1` without a Node.js traceback, policy picker EOF returns nonzero, and messaging mutation guards print the exact host command.
+  For more information, refer to [Recover and Rebuild Sandboxes](/user-guide/openclaw/manage-sandboxes/operate-sandboxes/recover-and-rebuild-sandboxes), [About Managed MCP Servers](/user-guide/openclaw/manage-sandboxes/mcp-servers/about-managed-mcp-servers), and the [NemoClaw CLI Commands Reference](/user-guide/openclaw/reference/commands).
+- The installer now requires an existing OpenShell executable to report a version after sandbox backups and before gateway or onboarding changes.
+  A version-tag install reports the requested tag, and a near-miss license response such as `y` receives an in-place hint without weakening exact `yes` acceptance.
+  WSL Express selects Windows-host Ollama only with Docker Desktop integration and otherwise configures WSL-local Ollama.
+  DGX Station preparation now distinguishes an active vLLM server from unrelated diagnostic processes.
+  For more information, refer to the [NemoClaw Quickstart with OpenClaw](/user-guide/openclaw/get-started/quickstart), [Update Sandboxes](/user-guide/openclaw/manage-sandboxes/operate-sandboxes/update-sandboxes), [Prepare Windows for NemoClaw](/user-guide/openclaw/get-started/additional-setup/windows-preparation), and [Prepare DGX Station to Install NemoClaw](/user-guide/openclaw/get-started/additional-setup/dgx-station-preparation).
+- The Ollama model menu now shows download size, required VRAM, and available or total GPU memory when known.
+  Resumed onboarding reports when a recorded reasoning setting takes precedence over `NEMOCLAW_REASONING`.
+  Passing the managed Hermes Dockerfile to `nemohermes onboard --from` now stages the repository root, and separate Hermes provider and model flags use the credential-resolving combined route.
+  For more information, refer to [Set Up Ollama](/user-guide/openclaw/inference/local-inference/set-up-ollama), [Configure Model Capabilities](/user-guide/openclaw/inference/manage-inference/configure-model-capabilities), [Install Hermes Plugins](/user-guide/hermes/manage-sandboxes/install-hermes-plugins), and the [NemoHermes CLI Commands Reference](/user-guide/hermes/reference/commands).
+- Both OpenClaw image paths now consume one authoritative production lock for OpenClaw `2026.7.1`.
+  Image assembly verifies the lock, registry integrity, installed package graph, and reviewed lifecycle before runtime exposure.
+  The OpenClaw graph replaces its affected `brace-expansion` and `fast-uri` resolutions.
+  All managed sandbox base images now install checksum-bound fixed Vim, jq, Oniguruma, and Expat packages and verify the reviewed Perl components.
+  Their bundled npm replaces its private `brace-expansion@5.0.7` copy with SRI-pinned `5.0.8`, and each image records a root-owned, read-only package inventory.
+  Hermes image assembly removes build-only caches and dependencies, reducing the measured final image by 624,394,525 bytes.
+  For more information, refer to [Architecture Details](/user-guide/openclaw/reference/architecture).
+- Release validation now runs approved E2E targets for fork PRs, retries one confirmed hosted-runner loss, and retries classified transient base-image pulls.
+  The required E2E observer retries transient GitHub reads after it revalidates the exact PR, head, and base identity, and redacts terminal network errors.
+  The reviewed npm audit gate retries only incomplete or timed-out scans within a fixed budget, evaluates complete vulnerability reports immediately, and fails closed when retries exhaust.
+  It reuses the reviewed Hermes production image, shards macOS and WSL tests, provisions pinned OpenShell on macOS shards, and enforces orchestration timeouts, credential boundaries, and artifact handoffs.