Add systems and HPC implementation guidance for coding agents

[yuecideng]

Description

Coding agents currently have HPC review guidance but lack an implementation workflow when computational costs matter. Add a conditional systems/HPC implementation skill, linked from AGENTS.md, solver scaffolding, pre-commit checks, and thin Claude/Copilot adapters.

The workflow starts from representative workloads and a cost model, then covers algorithm/backend choices, batching, bounded intermediate storage, buffer ownership, transfers, synchronization, and correctness/performance tradeoffs. Benchmark guidance now uses warmed repeated measurements with explicit synchronization, separates latency and memory passes, and distinguishes device-resident from end-to-end costs. Update solver kernel paths to compute/kinematics/_warp/ and fix literal newline escapes in the Markdown report example.

Dependencies: none. No production API or runtime behavior changes.

Type of change

• Documentation update

Validation

• black . using Black 26.3.1: 1,299 files unchanged.
• python docs/scripts/check_api_docs.py: 2,594/2,594 exports documented.
• Canonical skill validation and the new Claude adapter validation passed.
• context.py check and the focused context suite: 40 tests passed.
• All nine benchmark Python snippets parsed; helper smoke checks covered asynchronous completion timing, warmup exclusion, output lifetime, invalid repetition counts, a real CPU memory call, and multiline report generation.
• git diff --check passed. GPU hardware measurements were not run; this PR changes instructions and examples only.

Agent context

No topic or MAP changes are needed: implementation facts and subsystem ownership remain accurate. The reusable procedure belongs in a skill, while AGENTS.md retains the global trigger and entry point.

Screenshots

Not applicable.

Checklist

• I have run the black . command to format the code base.
• I reviewed affected documentation and agent context, updated it where needed, or explained why no update was needed.
• Public API documentation gate passed; no public API changes.
• Existing context tests and executable example checks passed; no new production tests are needed for this instructions-only change.
• No dependency changes are required.

[greptile-apps]

Confidence Score: 5/5

[Low impact] This documentation-only PR appears safe to merge.

Summary

Adds a conditional implementation workflow for performance-sensitive computation.

• Agents now have a shared workflow for costly computation.
• Benchmark instructions now measure timing and memory separately.
• New solver guidance points Warp kernels to the shared compute folder.

_{Reviews (1) · Last reviewed commit: "docs(agent): guide systems and HPC imple..." · Reviewed by Greptile}