[AIwork4me]

What & why

Wave-1 of the umbrella backport model (CONTRIBUTING §"Contributing between the two repos"): hardware-generic assets graduate from SenseNova-U1.5-ROCm-8060S (the Strix Halo lab, gfx1151) into this umbrella repo, which gains its first APU hardware profile. The 8060S repo stays fully independent — its quickstart, evidence corpus and releases are untouched; its alignment roadmap lands there separately (branch docs/upstream-alignment).

12 commits:

• patch 0003 — opt-in torch.compile/cudagraph safety (SENSENOVA_COMPILE* env knobs), renumbered from the 8060S 0002; auto-applied by the existing patches/*.patch loop, no setup-script change.
• scripts/gpu_monitor.py (VRAM+GTT sampling — GTT visibility is new capability) and scripts/rocm_check.py (env sanity incl. the APU GTT-spill alloc test), stdlib-only, torch imported only at runtime; 4 new offline unit tests.
• docs/porting.md + _CN — generic split of the 8060S porting notes (APU-specific depth moved to the profile).
• docs/hardware/strix-halo/ — first hardware profile (EN+CN) + hsa_fix.sh (byte-identical function body; now covered by CI shellcheck).
• docs/upstream/aotriton-54.md — tracking entry for the AOTriton hd=72 silent-wrong-output bug.
• governance — Code of Conduct, SECURITY.md, bug/validation issue templates, PR template (evidence gates), CONTRIBUTING cross-repo section (graduation criteria), README/README_CN hardware-coverage table.
• evidence plumbing — receipt.py/validate.sh now record hardware=<gfx> on every run.

Evidence

• patch 0003: git apply --check clean against the pinned upstream 76c32c2 (and sequentially after 0001+0002); real-GPU smoke on gfx1151 with SENSENOVA_COMPILE=1 (1024²×4 steps — [compile] enabled {'dynamic': False}, PNG saved, exit 0). V1–V8 speedup + Qwen-Image-Bench paired quality validation: 8060S evidence/speedup2 — verified on: gfx1151; gfx1100 receipt pending, compile stays opt-in.
• tools: real runs on gfx1151 — rocm_check.py: torch 2.12.0+rocm7.14.0, bf16 matmul ≈24.5 TFLOPS, --alloc-test 8 ok, exit 0; gpu_monitor.py: VRAM+GTT JSONL + peak summary verified live.
• docs: every number in porting/strix-halo/aotriton-54 traced to 8060S sources; gfx1151-only claims labeled verified on: gfx1151 with evidence links.
• aotriton-54: silent wrong output (not launch failure — that is the separate gfx1100 SDPA entry, cross-linked); open since 2024-11; our production repro + root-cause chain filed 2026-08-26.

Gates

• python3 -m pytest tests/ -q — 10 passed under CPU-only conditions. (The single local failure, test_fullstack_mode_logic, is host-environmental: it asserts on /opt/rocm full-stack preload on the authoring gfx1151 machine; reproduced identically on the pre-branch baseline; CI runners have no /opt/rocm and take the skip branch.)
• shellcheck scripts/*.sh scripts/lib/common.sh docs/hardware/strix-halo/hsa_fix.sh — clean.

Ported from (trailers in-commit): @ceefff1a (patch), @6046587f (tools ×2, porting docs, strix-halo profile), @d0af6efc @987015a1 @28e4f3af (aotriton-54).

Notes for reviewers:

• Commits were replayed via the Git Data API (git transport is TLS-broken from the authoring host). The branch tree is byte-identical to the locally reviewed branch (tree SHA db1cb531 verified on both sides); commit SHAs differ from the local originals only by timezone normalization (+0800 → UTC, same instants).
• Rebased onto e9c7974: README/README_CN contain both this PR's hardware-coverage section and that commit's cfg_interval retraction.
• Heads-up: CHANGELOG [Unreleased] still carries the older cfg_interval −31.9 % (0.7–1.0) entry that e9c7974's n=36 retraction supersedes — pre-existing on main, worth a separate small docs fix.
• docs/upstream/aotriton-54.md is EN-only, matching the four existing docs/upstream/ entries (bilingual waiver noted for the record).