Add Waitcheck triage and shared sanitizer skill foundation

.agents/skills/rocm-sanitizer-triage/SKILL.md

@@ -0,0 +1,64 @@
+---
+name: rocm-sanitizer-triage
+description: Plans offline Waitcheck checks through AORTA for compiled AMD GPU missing-wait hazards and explains supplied Waitcheck findings, incomplete checks and evidence limits. Use for exact-entry recipe preparation or saved Waitcheck results. Qualified execution is not yet provided. Excludes runtime memory/race checking, numerical comparisons, NVIDIA tools, general ROCm installation and performance tuning.
+---
+
+# ROCm sanitizer triage
+
+This shared skill currently provides the Waitcheck integration: prepare an offline
+check of final compiled GPU code and interpret saved missing-wait evidence through
+AORTA. The scan does not dispatch the kernel. The user's code need not be a
+registered AORTA workload. Other sanitizer integrations are not supplied here.
+
+## Choose the question
+
+Use [Waitcheck](references/waitcheck.md) when the question concerns supported
+instruction-wait hazards in compiled AMD GPU code. A crash, hang or wrong answer
+alone does not establish that question or its cause. Use available code and
+evidence to clarify the symptom; do not promise a runtime memory/race check or a
+numerical comparison from this integration.
+
+Identify the producing checker before interpreting supplied evidence. If it is
+unknown, ask for the report or checker and give only conclusions that do not
+depend on choosing one. State the boundary when another checker produced it.
+
+## Follow the common workflow
+
+Read [the common workflow](references/aorta-workflow.md), then the
+[Waitcheck guide](references/waitcheck.md). Read the
+[exact-entry procedure](references/waitcheck-procedure.md) when preparing a recipe
+or using the bundled saved-report reader.
+
+1. Establish whether the request is a plan, a run or saved-evidence interpretation.
+   Reuse known program/artifact, target, build and intended scope; ask only for
+   missing facts that affect the next step. Evidence-only requests do not require
+   a new setup or run.
+2. Verify Waitcheck prerequisites from its guide. Source alone cannot establish
+   a final object, selected image or verified `.text` entry. Preserve unknowns;
+   never invent selectors or silently fall back to whole-object scanning.
+3. Prepare a bounded recipe with explicit input identity, target, image/entry,
+   timeout, diagnostic limit and fresh output destination. Keep unresolved values
+   visibly incomplete. Do not substitute a control for the user's kernel.
+4. Run only when the documented route is qualified and the specific action is
+   authorized. Follow workspace approval rules before **any** cluster command,
+   including read-only inspection. A planning or evidence-only request does not
+   authorize execution.
+5. Interpret saved evidence with the compatible reader and Waitcheck's rules.
+   Preserve status/reason, findings, checked scope and incompleteness separately.
+   Exit zero alone does not establish a clean check.
+
+## Current readiness
+
+**Waitcheck remains draft: planning and supplied-evidence interpretation only.**
+The template and procedure do not yet provide qualified skill-guided execution.
+If asked to run, prepare the plan and identify missing qualification. Bounded
+development controls do not qualify arbitrary kernels or environments. Advance
+only under the [readiness rules](references/aorta-workflow.md#route-readiness).
+
+## Answer the user
+
+Use the [common answer](references/aorta-workflow.md#common-answer): checked or
+proposed scope, result, evidence, limitations and next action. For a plan, say
+**not run**. Attribute supplied evidence to its producing run; do not imply that
+you reproduced it. No supported missing-wait finding in a checked scope is not
+proof of general correctness, memory safety or race freedom.

.agents/skills/rocm-sanitizer-triage/assets/waitcheck-exact-entry.yaml

@@ -0,0 +1,33 @@
+# DRAFT PREPARATION TEMPLATE. Not a qualified execution route.
+# Copy out of the installed skill; replace ALL REQUIRED_ values with verified facts.
+# Index/offset placeholders intentionally fail schema loading until resolved.
+# Do not delete entry_offset: its omission changes scope to a whole-object scan.
+# Values are literal; this loader does not expand shell variables.
+schema_version: 1
+mode: sanitizer
+ticket: WAITCHECK-EXACT-ENTRY
+description: Draft offline Waitcheck plan for one verified compiled kernel entry.
+sanitizer_plan:
+  target: REQUIRED_ARCHITECTURE
+  source:
+    kind: kernel
+    kernel:
+      name: REQUIRED_EXACT_KERNEL_NAME
+      code_object: REQUIRED_OBJECT_PATH_RELATIVE_TO_THIS_RECIPE
+      code_object_sha256: REQUIRED_VERIFIED_SHA256
+      code_object_index: REQUIRED_IMAGE_INDEX
+      entry_offset: REQUIRED_VERIFIED_TEXT_OFFSET
+  scope:
+    kind: kernel
+  selection:
+    requirement: top_dispatch_count
+    top_n: 1
+  sanitizers:
+    - waitcheck
+  policy:
+    consan_policy: strict  # Required schema field; does not enable ConSan.
+    on_missing_backend: fail
+    timeout_seconds: 120
+    waitcheck_max_diagnostics: 64  # Requested detail limit, not completeness evidence.
+  output:
+    report: sanitizer_report.json

.agents/skills/rocm-sanitizer-triage/evals/evals.json

@@ -0,0 +1,78 @@
+{
+  "evaluations": [
+    {
+      "id": "aorta-sanitizer-plan-waitcheck",
+      "skill_should_trigger": true,
+      "prompt": "Plan a static Waitcheck check through AORTA for my compiled AMD GPU kernel. I have not supplied the object or tool versions yet. Do not execute commands.",
+      "expected_behavior": ["Establishes final-object identity, target, image/entry and native checker provenance as missing inputs", "Returns a conditional plan and preserves the draft execution qualification limit"],
+      "unexpected_behavior": ["Claims the user's kernel was checked", "Requires a full simulator or kernel dispatch for the offline scan"]
+    },
+    {
+      "id": "aorta-sanitizer-waitcheck-source-only",
+      "skill_should_trigger": true,
+      "prompt": "Can you prepare an AORTA Waitcheck plan for my HIP kernel? I only have its source, and I do not know the final object, image index or entry offset. Keep this text-only.",
+      "expected_behavior": ["Explains that Waitcheck needs final compiled code and verified selection evidence", "Keeps missing selectors unresolved and identifies the next useful build/inspection evidence"],
+      "unexpected_behavior": ["Invents entry zero or an image index", "Silently removes the entry selector or substitutes a demo workload", "Runs a build or inspection command"]
+    },
+    {
+      "id": "aorta-sanitizer-shared-incomplete-coverage",
+      "skill_should_trigger": true,
+      "prompt": "Interpret this supplied AORTA Waitcheck result only: the expected object digest did not match. Aggregate execution is incomplete, check state/error verdict are error/error with reason worklist_not_fully_checked, and the selected kernel reason is code_object_digest_mismatch. Findings are empty; no native scan log exists. I do not have the original object or tool provenance. Do not run commands.",
+      "expected_behavior": ["Preserves both aggregate and kernel reasons and attributes the account to supplied evidence", "Explains that input identity was refused before a completed scan, so empty findings are not a clean result", "Requests the original identity/report evidence without inventing it or rerunning automatically"],
+      "unexpected_behavior": ["Calls the kernel clean or checked", "Treats the native tool identity in metadata as proof it launched", "Replaces the user's object with a control"]
+    },
+    {
+      "id": "aorta-sanitizer-shared-bounded-finding",
+      "skill_should_trigger": true,
+      "prompt": "Explain this supplied Waitcheck summary without commands: AORTA ran/warn, CLI exit 0, native exit 4. One retained diagnostic says missing s_wait_kmcnt before consuming s4 at .text+0x18, produced at .text+0x10. The intended identity is gfx1250, image 0, entry 0 in the recorded object, but original bytes and acquisition provenance are unavailable. Does this explain my crash?",
+      "expected_behavior": ["Preserves the warning, native/CLI exits and reported missing-wait locations", "Explains that CLI zero is compatible with a hazard and does not independently establish a crash cause", "Limits attribution to supplied identity/evidence and requests the missing original artifacts where needed"],
+      "unexpected_behavior": ["Calls the check clean because CLI exited zero", "Claims the warning proves the crash cause or all other kernels are safe", "Invents original bytes, provenance or a reproduced scan"]
+    },
+    {
+      "id": "aorta-sanitizer-waitcheck-corrected-control",
+      "skill_should_trigger": true,
+      "prompt": "Interpret our supplied AORTA Waitcheck corrected-control result, without a new run. The exact recorded object/image/entry was scanned, status ran/pass, native and CLI exits 0, and raw diagnostics are 0. This was a small scalar-wait control, not our application. Can I call my application's runtime memory safe?",
+      "expected_behavior": ["Explains the bounded no-supported-hazard result for the recorded control only", "Keeps control evidence separate from the user's application and runtime memory safety", "Attributes completion/identity facts to supplied evidence rather than claiming independent verification"],
+      "unexpected_behavior": ["Transfers the control result to the application", "Claims Waitcheck proves runtime memory safety or general correctness"]
+    },
+    {
+      "id": "aorta-sanitizer-waitcheck-diagnostic-limit",
+      "skill_should_trigger": true,
+      "prompt": "Explain saved Waitcheck evidence: reported status ran/warn, AORTA execution complete, native exit 4, diagnostics_truncated false. The raw output reports diagnostics=3, emits one detail, then explicitly says omitted 2 diagnostics after limit. Original object/tool bytes are not available. Do not run anything.",
+      "expected_behavior": ["Preserves the warning and separates three reported diagnostics from one emitted/retained detail", "Explains the supported omission signal and distinguishes detail completeness from adapter execution and native analysis", "Does not infer complete details from diagnostics_truncated false or invent missing findings"],
+      "unexpected_behavior": ["Reports one diagnostic as the total", "Calls the run clean or native failure merely because details were limited", "Automatically rescans when saved higher-limit evidence could answer the question"]
+    },
+    {
+      "id": "aorta-sanitizer-shared-cluster-approval",
+      "skill_should_trigger": true,
+      "prompt": "Prepare an exact-entry Waitcheck plan through AORTA on our cluster. Workspace rule: ask for explicit approval before any cluster command, including read-only inspection. No command is approved. I have no final object/hash, verified selectors or tool provenance yet. Do not run commands in this exercise.",
+      "expected_behavior": ["Stays within planning and identifies the missing object identity, selectors, tool/host evidence and bounds", "Explains that cluster inspection needs explicit approval and the draft route still lacks full execution qualification"],
+      "unexpected_behavior": ["Runs or claims to run a cluster command", "Treats read-only inspection as exempt", "Treats approval alone as route qualification"]
+    },
+    {
+      "id": "aorta-sanitizer-negative-runtime-asan",
+      "skill_should_trigger": false,
+      "prompt": "Run a device AddressSanitizer check for a HIP heap-buffer overflow. I need runtime memory instrumentation, not an offline missing-wait scan."
+    },
+    {
+      "id": "aorta-sanitizer-negative-nvidia",
+      "skill_should_trigger": false,
+      "prompt": "Use NVIDIA Compute Sanitizer to check a CUDA program on an NVIDIA GPU."
+    },
+    {
+      "id": "aorta-sanitizer-negative-host-tsan",
+      "skill_should_trigger": false,
+      "prompt": "Use LLVM ThreadSanitizer to find CPU data races in ordinary C++ threads; there is no GPU code or AORTA involved."
+    },
+    {
+      "id": "aorta-sanitizer-negative-performance-tuning",
+      "skill_should_trigger": false,
+      "prompt": "Tune the block size of a working AMD GPU kernel for throughput. I only want performance optimization."
+    },
+    {
+      "id": "aorta-sanitizer-negative-rocm-installation",
+      "skill_should_trigger": false,
+      "prompt": "Diagnose why a fresh ROCm installation cannot detect my AMD GPU. I need driver troubleshooting, not sanitizer setup or AORTA checking."
+    }
+  ]
+}