test: record Qwen3.5 2B Q4_K_M physical acceptance

[daniele21]

Adds a dedicated host-owned physical acceptance lane for the exact curated Qwen3.5 2B Q4_K_M artifact already installed inside Harnex, and records the Samsung SM-A566B PASS captured on 2026-09-07.

Observable outcome:

• no Mac-side GGUF path or duplicate model copy;
• Harnex runs its existing PhoneTestController.runFullValidation() path against the selected private-ModelStore artifact;
• the runner captures the visible privacy-safe report, screenshot and bounded manifest through ADB;
• exact SHA-256 / architecture / quantization identity is enforced;
• PASS requires generation, cancellation, five repeated load/generate/unload PSS samples and physical thermal evidence.

Physical evidence recorded:

• device: Samsung SM-A566B, Android 16 / API 36 / arm64-v8a;
• model SHA-256: aaf42c8b7c3cab2bf3d69c355048d4a0ee9973d48f16c731c0520ee914699223;
• TTFT: 779 ms;
• cancellation: PASS;
• PSS samples: [251227, 239881, 251721, 252081, 252743] KB;
• PSS growth: 1,516 KB, below the existing 131,072 KB budget;
• thermal: 0 -> 0;
• validation duration: 25,303 ms;
• manifest SHA-256: 2e5e87749ea38a2bb9d309da43f9c112ee0bc93742c92b55e19b5de293095535.

Evidence boundary:

• outputTokens=1, therefore the reported decode tok/s is not promoted as a representative performance baseline;
• sustained-throughput benchmarking remains separate;
• this does not replace release-source identity, LOW_MEMORY, cross-model switching, or other release-critical real-environment gates.

[daniele21]

Superseded by #554, which reapplies the exact Qwen3.5 2B Q4_K_M physical evidence and host-owned collector on the current dev baseline after #549, together with the reconciled release ledger.