rdmsm4x-changelog-20260923-2301-HyperCPU-H3-GPU-and-system-telemetry

HyperCPU H3 — GPU and sysctl telemetry handback

Completed 2026-09-23 on rdmsm4x in the isolated handoff clone. Root: /Users/richh/dev/_handoff/hypercpu-gpu-sysctl-20260923-codex/apps/HyperCPU Branch: codex/gpu-sysctl-20260923; base: 106b689b80ce97a4645adfd08badfa16289c4e27. Production / App Store bound. Contract: LANES-H.md H3 and DEC-HYPERCPU-20260923-01. No push, installation, helper activation, canonical source changes, or Intel-host writes.

Separate deliverable commits

Commit	Deliverable and reason
`2ec4cff`	Common IOAccelerator per-device GPU telemetry on both architectures; additive optional `gpus` snapshot field and Apple Silicon utilization tile replace the unavailable primary clock tile.
`970312c`	Binary sysctl system sampler at existing snapshot cadence in both native and privileged providers; additive optional `systemStatistics`; task/thread keys accurately labelled as limits.
`370213d`	GPU and System chart tabs plus All integration; permanent series, full-opacity persistent legends, fixed 0–100 GPU utilization, window-retained growth-only memory/system axes, fixed first-observed maxfiles ceiling line, Space Blue palette.
`2aaa418`	CFBundleVersion 3 in Support/Info.plist, Info.plist and project.yml to identify the new artifact.

Evidence/documentation commit follows these four. All staging uses explicit paths.

Measurements and exact keys

Both architectures use public IOAccelerator PerformanceStatistics. No IOReport fallback was necessary for the requested GPU channels.

Architecture	Utilization	Memory
Apple Silicon M4 Max / rdmsm4x	`Device Utilization %`, `Renderer Utilization %`, `Tiler Utilization %`	`In use system memory`, `Alloc system memory` (bytes; unified memory, not VRAM)
Intel + AMD Vega20 / rdmpw3275m	`Device Utilization %`; `GPU Activity(%)` fallback	`inUseVidMemoryBytes` is present and used for VRAM used. Existing legacy discrete telemetry also uses `vramFreeBytes` for total=used+free. Renderer/tiler and Apple unified-memory counters remain unavailable.

GPU Core Utilization remains a compatibility fallback after the measured device/activity keys; it was not the source for these host captures. Missing/invalid counters remain unavailable and produce no zero/stale replacement samples. Real zero utilization is retained.

Identity: the live backend walks IOService parents to the nearest IOPCIDevice, grouping by its registry ID. Several services on one PCI device yield one GPU, never a sum/average. Pick the most complete PerformanceStatistics dictionary; break ties by accelerator registry ID. Distinct physical cards with identical IOClass remain distinct. AGX uses its accelerator registry ID when no PCI ancestor exists. IOClass/name fallback exists only for older/injected backends without registry identity.

AMD capture: eight current PerformanceStatistics records, four distinct accelerator registry IDs. The earlier count of sixteen includes PerformanceStatisticsAccum. Raw excerpts cannot establish parent PCI IDs; synthetic fixtures test PCI deduplication, and live Intel traversal remains unproven. Raw excerpts and recorded sysctl values: Tests/PowerGadgetCoreTests/Fixtures/H3/.

Exact system keys on both architectures:

kern.num_files vs kern.maxfiles
kern.num_vnodes vs kern.maxvnodes
kern.num_tasks, kern.num_threads
vm.loadavg (1/5/15, Darwin fixed-point loadavg structure)
vm.swapusage (total/used/free, Darwin xsw_usage structure)

The sampler uses sysctlbyname directly, not a subprocess, at each existing provider snapshot. No extra timer or entitlement. Read failures are unavailable with the OS error explanation. Integer width, missing/denied reads, invalid data, load scale, and recorded binary layouts are tested.

Semantic correction: XNU maps num_tasks to task_max and num_threads to thread_max. The chart says Task / Thread Limits; these are not current live counts. num_vnodes can exceed maxvnodes (the M4 capture does), so vnodes are not clamped. Source: https://github.com/apple-oss-distributions/xnu/blob/main/bsd/kern/kern_sysctl.c .

Validation

From the root above:

export CLANG_MODULE_CACHE_PATH=/private/tmp/hypercpu-h3/clang
export SWIFTPM_MODULECACHE_OVERRIDE=/private/tmp/hypercpu-h3/swiftpm
swift build && swift test
./scripts/build.sh

All exit 0 outside the execution sandbox (SwiftPM manifest sandboxing requires this). 253 tests = 210 Swift Testing in 22 suites + 43 XCTest; zero failures. Baseline 237; added 5 GPU + 5 sysctl + 6 chart/compatibility tests. Existing 15 fade tests and 8 permanent-domain tests pass. No test removals. The pre-existing unused baselinePacketData test warning remains; release source compiles with warnings-as-errors.

Full logs: docs/evidence/h3/swift-build.log, swift-test.log, release-build.log, artifact-verification.txt. GPU fixture test log also retained. Controls include real AMD VRAM, real Apple renderer/tiler/memory, distinct same-class PCI cards, duplicate services, real zeros, unavailable counters, malformed sysctls, history expiry, nonshrinking envelope, fixed ceiling retention and old snapshot decoding without the additive fields.

Built artifact

App: /Users/richh/dev/_handoff/hypercpu-gpu-sysctl-20260923-codex/apps/HyperCPU/dist/HyperCPU.app
ZIP: same directory, dist/HyperCPU.zip.
lipo -archs dist/HyperCPU.app/Contents/MacOS/HyperCPU: x86_64 arm64 (helper also Universal2).
Version: 1.0.0 (3); build script takes the bundle plist from Support/Info.plist.
Authority: Apple Development: Richard Doty (S65Q255HA8) → Apple Worldwide Developer Relations Certification Authority → Apple Root CA.
Team: ZU2882L4HT; hardened runtime; timestamp Sep 23 2026 22:58:21 EDT.
codesign --verify --deep --strict --verbose=2: valid on disk; satisfies its Designated Requirement, exit 0.
Executable SHA-256: f300935564b528b37c1da8d49ec5f6caf68699d0a3ddc2c8f8a4a5c87f044b4e.
Existing build uses macOS 15.0 minimum in the Intel Mach-O slice; unchanged by this lane.

Sandboxed codesign initially reported unavailable authority/invalid signature; the same artifact verifies successfully outside that sandbox. Final verification file contains the successful command outputs and individual exit codes.

Boundaries / not proven

No Intel executable ran or was installed; only authorized ioreg/sysctl OS measurements ran over SSH. Fixture coverage is not live Intel acceptance.
No GUI launch, screenshot, sustained chart-performance test or visual acceptance; lead owns candidate installation and all-six-Mac QA.
Default build.sh local Apple Development artifact; not a Developer ID notarized release or App Store sandboxed package. App Sandbox runtime access to these keys and App Store review acceptance remain unproven.
The old unused Sources/App Xcode target was not migrated; build.sh builds Sources/PowerGadgetApp + PowerGadgetCore + PowerGadgetUI. Its legacy GPUReader is not used by this delivered artifact.
GPU clock/residency remain unavailable where their own sources are unavailable. Utilization is never substituted into MHz or residency fields. Original four chart implementations/projections and menu bar strings retain their semantics.
No claim of physical GPU count from a live parent traversal on AMD, only four distinct captured accelerator IDs.

Persistence and rollback

SESSION-STATE.md and local PROJECTS.md updated. Notes/archive publication is recorded separately below. Rollback before integration: discard this isolated branch/artifact. After integration, revert the four feature/build commits in reverse order; this lane never changed installed apps or user data. Source baseline remains in Git at 106b689.

Notes publication verified 2026-09-23 23:01 EDT: Notes / rdmsm4x / rdmsm4x-changelog-20260923-2301-HyperCPU-H3-GPU-and-system-telemetry. Archive copies saved under ~/dev/LLM/Claude/changelogs/ and ~/dev/LLM/Codex/changelogs/. No owner action is needed for persistence.

Exact implementation files touched

Info.plist
Package.swift
Sources/PowerGadgetCore/Coordinator/TelemetryDispatchCoordinator.swift
Sources/PowerGadgetCore/Intel/IOAcceleratorGPUReader.swift
Sources/PowerGadgetCore/Mocks/MockTelemetryProvider.swift
Sources/PowerGadgetCore/Models/GPUDeviceMetrics.swift
Sources/PowerGadgetCore/Models/HardwareTelemetrySnapshot.swift
Sources/PowerGadgetCore/Security/PrivilegedHelperTelemetryProvider.swift
Sources/PowerGadgetCore/System/SystemStatisticsSampler.swift
Sources/PowerGadgetUI/Cards/AppleSiliconCard.swift
Sources/PowerGadgetUI/Charts/GPUAndSystemCharts.swift
Sources/PowerGadgetUI/Charts/SeriesPresence.swift
Sources/PowerGadgetUI/Dashboard/DashboardView.swift
Sources/PowerGadgetUI/Models/ChartDataModels.swift
Support/Info.plist
Tests/PowerGadgetCoreTests/Fixtures/H3/rdmpw3275m-ioreg.txt
Tests/PowerGadgetCoreTests/Fixtures/H3/rdmpw3275m-sysctl.txt
Tests/PowerGadgetCoreTests/Fixtures/H3/rdmsm4x-ioreg.txt
Tests/PowerGadgetCoreTests/Fixtures/H3/rdmsm4x-sysctl.txt
Tests/PowerGadgetCoreTests/GPUDeviceTelemetryTests.swift
Tests/PowerGadgetCoreTests/H3ChartTests.swift
Tests/PowerGadgetCoreTests/SystemStatisticsTests.swift
project.yml

Documentation: HANDBACK.md, SESSION-STATE.md, PROJECTS.md, Tests/PowerGadgetCoreTests/Fixtures/H3/README.md and docs/evidence/h3/.