[0.3] Add CPU-first host discovery and placement foundation

[pschildgen87-code]

Summary

This is the next Runtime 0.3 slice after the direct llama.cpp path. It establishes the host-resource contracts required before Runtime can make real backend/model activation decisions.

• adds quantum.runtime/host-profile/v1alpha1 Linux discovery for CPU vendor/model/features, physical/logical CPUs, SMT ratio, NUMA nodes, RAM/huge-page metadata, block/NVMe devices and DRM-visible AMD/NVIDIA/Intel accelerators
• discovery is fail-soft for optional sysfs data but validates core CPU/memory invariants before exposing the profile
• adds explicit bounded calibration with a worker-count sweep across physical-core and logical/SMT candidates, recording memory copy/read throughput and the best observed worker count
• calibration is explicit through POST /v1/host/calibrate; it does not consume resources automatically during daemon startup
• adds quantum.runtime/placement-plan/v1alpha1 with separate model-weight, MoE-expert, KV-cache, prefix-cache, projector, workspace and cold-tier classes
• implements the first CPU-first policy: CPU/RAM candidate first, GPU/hybrid only when the hot set does not fit and acceleration is explicitly allowed
• allows model/MoE weight splitting across RAM/VRAM as a pre-activation candidate while keeping cache/workspace classes whole-tier in this slice
• adds explicit NVMe placement only for declared cold bytes; active/hot state is never silently spilled to disk and capacity failures return capacity_exceeded
• adds /v1/host, /v1/host/calibrate and /v1/placement
• adds resource-control state for the last bounded calibration without mixing hardware discovery, calibration and placement responsibilities
• documents the current boundaries and Turin Tier-1 relationship in docs/HARDWARE-PLACEMENT.md
• bumps Runtime to 0.3.0-alpha.3

Validation already performed locally on the new packages

• gofmt
• go test ./... for hostprofile, calibration, placement, resourcecontrol
• tested discovery on Linux with CPU/NUMA/RAM/block-device data
• placement tests verify CPU-only remains preferred even when a GPU exists, hybrid is only considered after CPU capacity fails, explicit cold NVMe use, and no hot-state disk spilling

The repository PR CI remains the authoritative full-tree validation (gofmt, legal/systemd checks, version consistency, go vet, go test -race ./..., Runtime/installer builds).

Scope boundary

This slice does not yet auto-launch/reconfigure llama-server, persist calibration across restarts, set NUMA/thread affinity, estimate GGUF/KV sizes automatically, measure real-model prefill/decode throughput, dynamically query free VRAM, or claim a hybrid candidate is faster than CPU-only. Those require backend/model-specific conformance and the Turin hardware benchmark track.

Refs #8, #11, #12, #13.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
🔒 Security Review	✅ Completed 2026-09-05T21:59:10.277366Z	f728c5b	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.