UX: wizard-first launch + cross-platform docs + zero-config (adapted from club-3090)

[Sandermage]

What & why

Newcomer onboarding was expert-only: to reach a working curl you had to know a preset key, a pin, a models dir, and that the engine only runs on Linux+CUDA. This branch adds a zero-decision front door on top of the existing expert surface — adapting club-3090's per-hardware front-door pattern to our stack. Nothing was removed: sndr launch <preset>, explicit --pin, all existing verbs and flags keep working unchanged.

What got simpler (before -> after)

Decision a new user faced	Before	After
Which preset?	pick a key from the registry by hand	sndr quickstart auto-detects GPUs, projects per-card VRAM, picks a fitting preset (explicit > pinned-default > top-fit)
Boot it	know sndr launch <preset> + flags	sndr quickstart delegates to sndr up; can save the pick as default
No Linux GPU (Mac/Windows)?	dead end	sndr remote setup <url> one-liner -> client mode against a remote rig; per-OS docs (RUN_ON_MAC / RUN_ON_WINDOWS_WSL)
Config	hand-write env	.env.example scaffold; cp .env.example .env boots on built-in defaults; host.yaml auto-init on first run
Install	one profile	install.sh auto-selects client vs full-stack by CUDA presence

Files

Added (nothing removed):

• CLI: sndr/cli/user_prefs.py, sndr/cli/commands/quickstart.py, sndr/cli/commands/remote.py
• Config: .env.example, sndr/model_configs/host_autoinit.py, scripts/start_sndr_client.sh
• Docs: docs/RUN_ON_LINUX.md, docs/RUN_ON_MAC.md, docs/RUN_ON_WINDOWS_WSL.md, docs/REMOTE_ENGINE.md
• Tests: 9 new test files (user_prefs, quickstart, chat remote-env, up remote/default, remote setup, env_example, host_autoinit, client_profile, situation_docs)

Edited (additive/corrective only): sndr/cli/commands/{__init__,up,run,chat}.py, sndr/model_configs/host.py, install.sh, scripts/security_scan.py, .gitignore, README.md, docs/{GETTING_STARTED,QUICKSTART,README}.md, mkdocs.yml, tests/unit/scripts/test_security_scan.py.

Verification

• Full suite: 15065 passed, 586 skipped, 1 xfailed, 0 failed.
• make_evidence 65/65 green; check_doc_sync --strict clean (329 patches); test_doc_preset_keys 2/2.
• audit-public-paths clean (fixed a literal LAN IP in remote.py examples -> <your-host>).
• 0-new-ruff whole-file on all changed .py (ruff 0.15.11).
• Server-tested on the rig, PROD vllm-35b-dev748 untouched: sndr quickstart --dry-run --no-input auto-detected OS + 2x A5000 via nvidia-smi and reached the boot decision with no manual config; sndr remote setup probed live PROD :8102 from the Mac (reachable); install.sh --dry-run (full-stack + client) scaffolded .env and host.yaml auto-init.

Known follow-up (not blocking)

• Zero-decision default resolves to prod-qwen3.6-35b-multiconc (throughput-tuned, peak ~100% per-card VRAM) while README documents prod-qwen3.6-35b-balanced as the lone-user default. Reconcile the ranking bias or the doc claim in a fast-follow.

[gemini-code-assist]

Code Review

This pull request introduces a 'client mode' that enables users on Mac or Windows to drive a remote Linux rig as their engine, effectively decoupling the CLI/GUI from the local hardware. Key additions include the sndr quickstart command for automated setup, sndr remote setup for configuring remote connections, and updated documentation for cross-platform deployment. The review feedback highlights necessary improvements for Python 3.10 compatibility, URL validation logic, VRAM projection accuracy, and installer script robustness.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Since Python 3.10 is explicitly supported by the installer, importing tomllib directly will raise a ModuleNotFoundError on Python 3.10 because tomllib was introduced in Python 3.11. To maintain compatibility, use a fallback import to tomli (or set it to None to handle it gracefully with a fallback parser).

Suggested change

	import tomllib
	try:
	import tomllib
	except ImportError:
	try:
	import tomli as tomllib
	except ImportError:
	tomllib = None

[gemini-code-assist]

To support Python 3.10 without requiring external dependencies, provide a simple fallback parser when tomllib is unavailable. This also avoids an AttributeError when referencing tomllib.TOMLDecodeError if tomllib is None.

def _read() -> dict[str, Any]:
    path = _prefs_path()
    if not path.is_file():
        return {}
    if tomllib is not None:
        try:
            with path.open("rb") as handle:
                return tomllib.load(handle)
        except (OSError, Exception):
            return {}
    # Fallback parser for Python 3.10 when tomli/tomllib are unavailable
    try:
        data: dict[str, Any] = {}
        current: dict[str, Any] = {}
        for line in path.read_text(encoding="utf-8").splitlines():
            line = line.strip()
            if not line or line.startswith("#"):
                continue
            if line.startswith("[") and line.endswith("]"):
                sec = line[1:-1].strip()
                current = {}
                data[sec] = current
            elif "=" in line:
                k, v = line.split("=", 1)
                current[k.strip()] = v.strip().strip('"').strip("'")
        return data
    except OSError:
        return {}

[gemini-code-assist]

The check if parsed.path.rstrip("/") evaluates to False when the path is empty or just / (e.g., http://localhost:8102 or http://localhost:8102/). This allows URLs without the required /v1 suffix to pass validation, which will fail later during API requests. Enforce that the path must end with /v1 by removing the falsy check.

Suggested change

	if parsed.path.rstrip("/") and not parsed.path.rstrip("/").endswith("/v1"):
	return None
	if not parsed.path.rstrip("/").endswith("/v1"):

[gemini-code-assist]

The gpu_count is hardcoded to 1 in ProjectorRig. For multi-GPU rigs, this will calculate the VRAM projection incorrectly as if only a single GPU is present. Use the detected rig.gpu_count instead.

Suggested change

	kp.ProjectorRig(vram_gib_per_card=vram_gib, gpu_count=1, name=rig.source),
	kp.ProjectorRig(vram_gib_per_card=vram_gib, gpu_count=getattr(rig, "gpu_count", 1), name=rig.source),

[gemini-code-assist]

When the installer is run via a piped stream (e.g., curl -sSL ... | bash), ${BASH_SOURCE[0]} is empty/unset, and $0 defaults to the shell executable path (like /bin/bash or /usr/bin/bash). This causes dirname "$0" to resolve to /bin or /usr/bin, which adds them to PYTHONPATH and searches them for .env.example. Avoid falling back to $0 when ${BASH_SOURCE[0]} is not set.

Suggested change

	SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]:-$0}")" 2>/dev/null && pwd 2>/dev/null || echo "$PWD")"
	SCRIPT_DIR="$PWD"
	if [ -n "${BASH_SOURCE[0]:-}" ]; then
	SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" 2>/dev/null && pwd 2>/dev/null || echo "$PWD")"
	fi

[gemini-code-assist]

The pattern [A-Za-z_]* only validates the first character of the key. If the key contains invalid characters (like hyphens or spaces) later in the string, it will pass this check but cause a syntax or runtime error during indirect expansion ${!key} or export. Ensure the entire key contains only valid alphanumeric characters and underscores.

Suggested change

	case "$key" in [A-Za-z_]*) : ;; *) continue ;; esac
	case "$key" in [A-Za-z_]*) : ;; *) continue ;; esac
	case "$key" in *[!A-Za-z0-9_]*) continue ;; esac