feat(skills): add coding-guidelines skill and consolidate PR guidance

[demandal25]

Summary

Two skill/docs changes: adds a new coding-guidelines skill, and consolidates the project's PR mechanics into the pr-workflow skill (moving them out of CLAUDE.md) so they live in one place and are followed consistently.

What changed

• .claude/skills/coding-guidelines/SKILL.md — new skill: YAML frontmatter (name / description) plus four numbered sections — Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. Mirrors the existing plan-review skill so the two read consistently.
• .claude/skills/pr-workflow/SKILL.md — absorbs the branch-naming and ask-before-pushing rules from CLAUDE.md, and gains PR body-formatting rules: do not hard-wrap PR bodies (GitHub renders a single newline as <br>), and do not append a "Generated with Claude Code" footer.
• CLAUDE.md — drops the now-relocated branch-naming and ask-before-pushing sections, keeping it lean and pointing at the pr-workflow skill.

Test plan

• Skill/docs-only change — no Python code paths touched, so there are no relevant pytests to run.
• pre-commit run --files on all three changed files — all applicable hooks pass (markdownlint included). These scoped runs cover the branch's entire diff; a full pre-commit run -a is impractically slow over the whole repo and not meaningful for markdown-only changes.

[Copilot]

Pull request overview

This PR expands ROCm batch-decode support by enabling opt-in CUDA-graph capture for the AITER batch-decode path (explicit backend="aiter" with a capture-at-max contract), adds tests + benchmarks for graph replay correctness/perf, and updates documentation/changelog accordingly. Note: the current PR title/description state this is only adding a coding-workflow skill, which does not match the actual diff scope.

Changes:

• Enable AITER batch-decode under CUDA-graph capture for explicit backend="aiter", replacing the prior hard error with a one-time warning describing the capture-at-max constraint.
• Add ROCm test coverage for AITER graph replay and add a benchmark script comparing graph replay behavior/performance.
• Update README/CHANGELOG for the new CUDA-graph/AITER behavior, and add the new .claude coding-workflow skill document.

Reviewed changes

Copilot reviewed 1 out of 1 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
flashinfer/decode_rocm.py	Allows backend="aiter" with CUDA-graph enabled (warns about capture-at-max constraint instead of raising).
tests/rocm_tests/test_batch_decode_aiter_hip.py	Removes the “graph capture rejected” test and adds a positive graph replay correctness test for AITER.
benchmarks/rocm_benchmarks/bench_decode_graph_hip.py	Adds an acceptance/timing script for AITER decode under graph capture + replay.
benchmarks/rocm_benchmarks/bench_batch_decode_hip.py	Adds/updates eager decode benchmark; docstring needs alignment with new graph behavior.
README.md	Documents the new opt-in AITER-under-graph behavior and clarifies auto routing under capture.
CHANGELOG.md	Adds a release note entry describing AITER CUDA-graph capture support and associated tests/benchmarks.
.claude/skills/coding-workflow/SKILL.md	Adds the new coding-workflow skill content (workflow guidance).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (1)

.claude/skills/pr-workflow/SKILL.md:209

• This note says the hard-wrap rule “mirrors the ## PR / issue body formatting rule in CLAUDE.md”, but that section was removed from CLAUDE.md in this PR, so the reference is now stale/inaccurate.

  to separate blocks. (Mirrors the `## PR / issue body formatting` rule in
  CLAUDE.md; note this is the opposite of how the repo's `.md`/`.py` source
  files — including this skill — are wrapped.)