docs/test: reconcile physical release evidence

[daniele21]

Outcome

Reconcile the physical Play and Qwen3.5 2B evidence against the current dev baseline after release-identity PR #549 merged.

This supersedes stale evidence PRs #551 and #553 without weakening any release gate.

Source identity

• base: dev@2ff88f05b5b9e0e2b62f3686a3bbb3aaa9ddbb53;
• candidate HEAD: bd994089508ff049c7de3d5752543e4e515da90a;
• candidate tree: 5641f4b566a1403487254d96cb988c7196983e0c.

Physical evidence recorded

Play Internal topology

The actual Play-delivered Harnex/RedactGuard builds passed:

• RedactGuard installed first;
• Harnex installed later without RedactGuard reinstall;
• PENDING observation;
• explicit Harnex authorization;
• Connect / Disconnect / Reconnect;
• real consumer inference.

Google Play Console reports the same Play App Signing SHA-256 identity for both current applications: D6:2D:3C:C8:51:D5:72:05:C3:42:C1:7F:86:26:40:58:E3:FE:29:6A:AE:1B:0E:43:FD:AC:58:82:24:44:1A:BD.

This closes physical install-order/authorization/connectivity for the current Play topology. It does not claim physical distinct-signer Play evidence; deterministic release-identity E2E remains the proof for that security boundary.

Qwen3.5 2B Q4_K_M

Samsung SM-A566B, Android 16/API 36/arm64-v8a, exact curated model digest aaf42c8b7c3cab2bf3d69c355048d4a0ee9973d48f16c731c0520ee914699223:

• host-owned production Android/JNI/backend generation PASS;
• active cancellation PASS;
• five repeated load/generate/unload cycles PASS;
• PSS samples [251227, 239881, 251721, 252081, 252743] KB;
• PSS growth 1,516 KB;
• thermal 0 -> 0;
• TTFT 779 ms;
• full validation 25,303 ms;
• evidence manifest SHA-256 2e5e87749ea38a2bb9d309da43f9c112ee0bc93742c92b55e19b5de293095535.

The sanity generation emitted one output token, so its reported decode tok/s is explicitly not promoted as representative sustained-throughput evidence.

Repository truth updated

• ADR 0018 now distinguishes intended/distinct-signer authorization from the current same-signer Play distribution reality.
• docs/current-state.md records both physical runs and replaces the stale “Play retest pending” blocker.
• docs/roadmap.md no longer lists the completed Play smoke as pending P0 evidence.
• Harness 0.5 checklist records these as partial physical evidence, closes only the already-integrated repo-side Overview/Diagnostics/Settings state/effect realignment, and leaves exact-release-candidate gates open.
• Adds a reusable host-owned 2B Q4_K_M acceptance collector that does not copy GGUF files from the Mac or reinstall Harnex.

Remaining release boundary

This PR does not declare Harness 0.5 release-ready. Exact-release-source/artifact binding, repository ruleset protection, navigation/process restoration, warm-idle TTL/RAM controls, representative responsive/accessibility evidence, applicable SR-6 physical lanes, LOW_MEMORY/cross-model lifecycle evidence, sustained performance/resource evidence where claimed, and 4B qualification remain open unless separately closed by exact-candidate evidence.

[daniele21]

/preflight auto

[daniele21]

/preflight auto

[daniele21]

/preflight auto

[daniele21]

/preflight auto