test(llup): replay automated v0.3 qualification evidence

[daniele21]

Scope

Stacked LLUP-60A evidence-only replay on top of frozen runtime candidate PR #490.

• runtime source: 59af48313b450d9cff13c7f43458c2e5e6560374;
• evidence HEAD: 2e213ad465bef56b724708e9b2b641bba0de3be5;
• base: agent/llup-v0-3-upgrade-rebased@59af48313b450d9cff13c7f43458c2e5e6560374.

The diff is intentionally restricted to three evidence/governance files:

• .github/workflows/llup60-automated-replay.yml;
• docs/workstreams/llup-v0-3-automated-replay.md;
• docs/workstreams/README.md for canonical active-workstream routing.

No runtime, backend, JNI, model-policy or product code is changed.

Refreshed qualification identities

The candidate was refreshed after dev gained the canonical Android AAB packaging command. Runtime behavior did not change, but exact build/package identity did, so pre-refresh LLUP-60 evidence is retained only as historical provenance.

Frozen pair:

• control runtime: dev@80164329bbc41a00b75721e3d0524294c03fdb56, llama.cpp b9637 / aedb2a5e9ca3d4064148bbb919e0ddc0c1b70ab3;
• candidate runtime: 59af48313b450d9cff13c7f43458c2e5e6560374, llama.cpp v0.3.0 / c1d0e7a004015f23bc0233470b747b596f29b264.

LLUP-60A automated replay — PASS

Authoritative exact-head evidence on 2e213ad465bef56b724708e9b2b641bba0de3be5:

• LLUP-60 automated replay 33335263956: SUCCESS;
• pull-request automatic Validate 33335266468: SUCCESS;
• repository-owned requested /preflight strong run 33335275065: SUCCESS.

The repository selector escalated the requested preflight to FULL with reason unknown or repository-wide executable scope, because this evidence PR changes a GitHub Actions workflow. Effective selector:
profile=full android=true native=false packaging=false modules=all.

This escalation is accepted rather than manually downgraded.

Replay lanes passed:

• evidence-only scope and exact runtime-parent provenance;
• authoritative pin/submodule/backend/Q35 identity + regression fixture;
• host-native backend ownership/API suite;
• Q35 model-profile + runtime-core JVM contracts;
• evaluation contracts across contracts/datasets/evaluators/engine/runtime-adapter/comparison/persistence;
• observability contracts + benchmark-engine unit contracts;
• exact Qwen3.5 0.8B and 2B GGUF download, digest/size verification and candidate llama-simple load/tokenize/generate smoke;
• machine-readable replay manifest with runtime source SHA separated from evidence-harness SHA and promotion explicitly disabled while physical evidence is pending.

Published exact-head artifacts:

• llup60-provenance;
• llup60-qwen35-host-compatibility;
• llup60-automated-replay-manifest.

Classify LLUP-60A = AUTOMATED_PREFLIGHT_CONFIRMED.

LLUP-50 deterministic preparation — PASS

Physical tooling owner PR #491 is refreshed at cefb893ae95cfd95339de4a24a955a652f0011e6 and its exact-head Validate run 33334618813 is green.

Frozen evidence refs:

• control: evidence/llup50-control@fcbefc7cd9af84de570da96d039582175dd1700b;
• candidate: evidence/llup50-candidate@a2a050d9551db541bb4c6b152cba8623c782164d.

Repository-owned package run 33334957429: SUCCESS. It produced exact-ref control, candidate and candidate-runtime app/test APK pairs plus source-revision/SHA-256 manifests.

Therefore there is no remaining deterministic package/tooling blocker before device execution.

Execution boundary

All remaining evidence is genuinely REAL_ENVIRONMENT:

• LLUP-50 same-device control/candidate A/B;
• physical LLRT KV-cache and evaluation-batch evidence;
• device-only Q35 load/performance/PSS/memory/thermal/cancellation/recovery/lifecycle evidence;
• representative-device resident-count, warm-idle and prepare-after-release recovery;
• output-quality replay whose canonical execution path genuinely requires representative Android runtime evidence.

Global LLUP-60 is not complete until the required device-only LLUP-60B evidence passes.

Current workstream classification

All required deterministic automated gates for the frozen candidate/base and LLUP-50 package preparation are green on exact identities.

Classify the LLUP workstream boundary as WAITING_REAL_ENVIRONMENT.

This does not authorize promotion:

• PR build(llup): migrate llama.cpp candidate to v0.3.0 #490 remains open and unmerged;
• this evidence PR remains draft/unmerged so the frozen runtime candidate SHA does not move;
• PR test(llup): add exact-APK LLUP-50 physical evidence runner #491 remains draft/unmerged while the physical A/B uses its frozen tooling;
• dev remains on the b9637 control baseline;
• LLUP-70 remains blocked on the physical evidence block;
• MRES-10 remains blocked by LLUP-70.

After LLUP-50 + LLUP-60B pass, re-check exact candidate/base freshness and FULL promotion requirements, then make the explicit LLUP-70 promote/reject decision.

[daniele21]

/preflight strong

[daniele21]

/preflight strong

[daniele21]

/preflight strong

[daniele21]

/preflight strong

[daniele21]

/preflight strong

[daniele21]

/preflight strong