Add bounded external KV-Ground compatibility smoke

[mark-e-deyoung]

Purpose

Implements the public-safe harness portion of #39 so opportunistic third-party/shared GPU capacity can falsify one concrete runtime assumption without becoming an MVP dependency.

Exact experiment

The harness pins:

• model: vocaela/KV-Ground-8B-BaseGuiOwl1.5-0315;
• exact corrected-weight revision: fe7563292bb52ab6c235fc3c87157e6a14017479;
• generated synthetic 1000x1000 GUI fixture;
• instruction: Click the TARGET button.

It tests the legacy-equivalent AutoModelForCausalLM path first. If that fails, the same environment can run the public evaluation repo's AutoModelForImageTextToText path. This distinguishes a bounded legacy-loader incompatibility from a broader environment/model failure without silently changing desktop-ui-cv #35.

Included

• experiments/kv_ground_external_smoke.py — portable contract-only/GPU smoke harness with canonical machine-readable receipts;
• experiments/tests/test_issue_39_external_smoke.py — stdlib-only contract/parser/repeatability tests;
• .github/workflows/issue-39-external-smoke-contract.yml — public Actions validation only; no model/GPU/provider access;
• docs/experiments/issue-39-external-kv-ground-smoke.md — execution order, interpretation, safety and non-qualification boundary.

Evidence boundary

The receipt may contain public/synthetic model output and environment facts. It intentionally has no private screenshots, target inventory, provider credentials, host traceback, desktop-ui-cv source, or WineBot material.

This PR does not establish:

• target Windows compatibility;
• private holdout quality;
• target VRAM/load/p95 latency;
• a Foundry qualification subject;
• an external GPU service abstraction.

The actual GPU execution remains opportunistic. If platform friction exceeds uncertainty reduction, #39 can stop without blocking the authoritative path.

Critical-path relationship

This remains parallel and nonblocking to:

exact identity -> exact provider-native materialization -> frozen request/target/envelope -> desktop-ui-cv #35 local handle -> private holdout + target resource measurement -> qualification receipt -> deterministic repeat -> one-material-change falsification -> GO/NARROW/KILL

Verification

The PR's public workflow validates the harness contract and byte-repeatable --contract-only output using only Python 3.12 and the standard library. It deliberately does not download model weights or request GPU capacity.

Stacked on the current Foundry control realignment branch from draft PR #22; retarget only when prerequisite control state lands.

Related: #18, #39; SemperSupra/desktop-ui-cv#35.