[codex] Sync Torch-OSQP release-readiness package

[Ztang-Yit-Xiaang]

Summary

• Sync the Torch-OSQP release-readiness package and documentation handoff.
• Add docs/plans/ as the decision-complete roadmap and phase-plan package.
• Update the main pipeline and completion audit to reference the new status source and keep accelerator promotion gated.
• Record fork-local CPU validation evidence for the current PR head and updated fork main.

Validation

Documentation/release-package checks:

• git diff --cached --check
• local docs/plans link check: link_problems=[]
• placeholder scan: no unresolved template tokens
• Mermaid/unrelated-content scan: no matches
• conservative backend scan: CUDA remains unpromoted; ROCm/MPS remain unclaimed
• PDF was regenerated and smoke-checked earlier: 19 pages with TOC, support matrix, risk table, roadmap, decision log, docs/plans/roadmap.md, and gated accelerator status

Fork validation evidence:

• PR head 88bfd164443d3804f57a3f0f1221c24f8daf0a9c core workflow passed: 28831049849
• PR head nightly CPU stability passed: 28831240981, 1800 CPU cases, zero release-gate failures
• Fork PR Ztang-Yit-Xiaang/PyGRANSO#2 merged the evidence commit into fork main: merge commit ffa9291a6a10ee6c4eec1025963512e6ca88b059
• Updated fork main core workflow passed: 28831702081
• Updated fork main nightly CPU stability passed: 28831873901, 1800 CPU cases, zero release-gate failures

Release notes

This PR is documentation/release-package stabilization only. It does not change solver source code and does not promote CUDA, ROCm, or MPS support.

Remaining blockers

• Upstream review/merge permission is still required for this PR.
• Register/merge default-branch workflows upstream before relying on upstream scheduled/nightly or manual hardware dispatch.
• Keep CUDA unpromoted until representative hardware passes correctness and the <=5x performance gate.
• Obtain ROCm and MPS runners before making either support claim.