Add: bench env knobs and unified commit/PR message conventions

[zhangqi-chen]

Summary

Benchmark knobs in golden/runner.py:

• PYPTO_BENCH_RAW prints every measured dispatch's Effective sample per rank — for L2, L3, and the L3 flatten fallback where the per-round summaries are least trustworthy; off by default
• PYPTO_BENCH_ROUNDS / PYPTO_BENCH_WARMUP override the hardcoded 100 / 5 loop sizes, which do not fit a long prefill or a multi-card L3 run; a malformed value warns and falls back instead of failing the run
• The resident L3 path clamps warmup to >= 1 via _resident_loop_sizes: it spends warmup[0] on the validation dispatch, so warmup=0 would emit rounds+1 dispatches and lose per-round segmentation
• Drop device_wall from every report (headline, per-rank, L3 detail); Effective is the metric consumed downstream. Daily CI sets none of the new vars, so its 100 / 5 baseline and the effective_us .*mean= number it collects are unchanged
• Unit tests for the knobs plus a guard that exactly one report line matches daily CI's collector; tests/golden/conftest.py clears the PYPTO_BENCH* vars, which previously failed 22 unrelated tests when exported

Benchmark documentation:

• docs/performance-tuning.md gains a "Measuring" section ahead of Part 1: enabling the timed loop, what Effective is, the mean= field daily CI collects, the real-device / SIMPLER_PROFILING requirements, reading the multi-card per-rank and fallback_flattened output, and the four PYPTO_BENCH* knobs
• docs/compile-runtime-workflow.md gains phase 4b — the loop runs after the correctness dispatch and before validation, and was undocumented

Commit / PR conventions in .claude/skills/:

• git-commit, github-pr and fix-pr rewritten against one shared convention: the seven commit types, a body derived only from the staged diff or BASE_REF..HEAD, ## Summary as the sole PR section, and no AI co-author footers
• git-commit gains the type table with per-type usage, the rule that a Perf: body states measured before -> after plus the configuration, and an explicit no-bypass rule for pre-commit hooks
• github-pr replaces the prose walk-through with the state table (effectively-on-main x uncommitted), the commit-type-to-branch-prefix table, and --force-with-lease after a rebase
• fix-pr drops the ASCII flow diagram for numbered steps with a 5-iteration cap

Documentation cleanup:

• Remove docs/investigations/2026-07-qwen3-fused-attention-sync-barrier.md; cce-extern-kernel-guide.md keeps the conclusion (the C220 MTE3/TSTORE-to-MTE2 path was validated over 40 layers without an extra DDR barrier) and cites perf(qwen3-14b decode): fold RoPE into the CANN FAI attention extern #796 in place of the deleted file

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[coderabbitai]

Important

Review skipped

Auto incremental reviews are disabled on this repository.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: a555de53-b1e8-43fd-8d66-dc30177abefd

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

📝 Walkthrough

Walkthrough

The PR rewrites commit, GitHub PR, and iterative fix workflows, and updates golden benchmarking with configurable loop sizes, optional raw timing output, revised L3 reports, and tests covering the new behavior.

Changes

Workflow guidance

Layer / File(s)	Summary
Commit and PR creation workflow .claude/skills/git-commit/SKILL.md, .claude/skills/github-pr/SKILL.md	Commit checks, branch naming, rebasing, authentication, and PR title/body rules are reorganized into explicit numbered workflows.
Iterative PR fixing .claude/skills/fix-pr/SKILL.md	PR input handling, review-thread and CI detection, issue classification, consent, commit folding, thread resolution, polling, and error handling are rewritten for an iterative five-pass workflow.

Benchmark reporting

Layer / File(s)	Summary
Benchmark sizing and reporting golden/runner.py	Benchmark rounds and warmup values can be configured through environment variables; raw per-dispatch samples are optionally reported, and L3 output omits device-wall metrics.
Benchmark path integration golden/runner.py	L2, non-resident L3, resident L3, and JIT paths use dynamic loop sizes and invoke raw reporting.
Benchmark validation tests/golden/conftest.py, tests/golden/test_runner.py	Tests isolate benchmark environment variables and cover loop-size defaults, overrides, invalid values, resident warmup behavior, and report formatting.

Estimated code review effort: 3 (Moderate) | ~25 minutes

Possibly related PRs

• hw-native-sys/pypto-lib#764: Related L3 benchmark loop, timing, and reporting changes.
• hw-native-sys/pypto-lib#778: Related L3 effective timing and CI-oriented reporter changes.

Poem

A rabbit reviews the loops with care,
Raw timing samples hop through the air.
PR threads resolve, commits neatly fold,
Warmups adjust as benchmarks are told.
“Green checks!” I cheer, with a twitch of my nose—
New guidance and reports in orderly rows.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title matches the main changes: benchmark env knobs and shared commit/PR message conventions.
Description check	✅ Passed	The description is directly related and summarizes the benchmark and workflow documentation changes.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 6

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

.claude/skills/fix-pr/SKILL.md (1)

28-51: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Use gh api graphql’s built-in pagination

The current instruction still asks the caller to rerun the query with a cursor and manually accumulate pages, but a second rerun still writes only the next page to /tmp/threads.json. Use gh api graphql --paginate --slurp with the query’s $endCursor: String variable so every page is requested and combined in one output.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.claude/skills/fix-pr/SKILL.md around lines 28 - 51, Update the GraphQL
command in the review-thread retrieval instructions to use gh api graphql
--paginate --slurp instead of manual cursor reruns and accumulation. Add the
pagination variable expected by gh, use it in reviewThreads, and remove the
guidance to rerun with -F cursor or manually combine pages; preserve the
existing /tmp/threads.json output and subsequent gh pr checks command.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In @.claude/skills/fix-pr/SKILL.md:
- Around line 104-117: Update the staging instructions in the “Fix and fold into
the PR's commit” section to remove git add -A and stage only the affected file
paths, explicitly including any intended new files. Preserve the existing fetch,
base-reference, and commit-count commands unchanged.
- Around line 141-149: Update the “Resolve comment threads” instructions to use
{owner} and {repo} placeholders in the gh api repository endpoint instead of
literal :owner and :repo segments. Ensure the surrounding guidance provides
enough repository context for GitHub CLI to resolve those placeholders, while
preserving the existing reply and resolve workflow.

In @.claude/skills/git-commit/SKILL.md:
- Around line 68-75: The commit-body guidance in the staged-diff verification
section conflicts with the Perf body requirement. Update the rule to require a
reproducible benchmark command and configuration for measured before/after
values, or explicitly permit external measurements only when their reproduction
details are included; apply the same clarification to the corresponding guidance
around the Perf format.

In @.claude/skills/github-pr/SKILL.md:
- Around line 22-44: Update the effectively-main branch flow in the github-pr
instructions to define commit subject generation before branch creation. Derive
the commit type and slug from the staged diff, then use that subject to select
the branch prefix and construct the branch name before invoking /git-commit;
preserve the existing naming rules and subsequent commit/PR actions.
- Around line 55-63: Require explicit human confirmation immediately before the
force-push step in both workflows: update .claude/skills/github-pr/SKILL.md
lines 55-63 around the rebase/push instructions, and
.claude/skills/fix-pr/SKILL.md lines 134-139 around its rewritten-commit push,
while leaving non-force pushes unchanged.

In `@tests/golden/test_runner.py`:
- Around line 1477-1481: Update the _FakeStats test double so mutable list
attributes are created per instance rather than shared at class scope,
addressing RUF012. Initialize the relevant invocation, dispatch, and wall-time
collections in _FakeStats.__init__ while preserving their existing defaults and
behavior.

---

Outside diff comments:
In @.claude/skills/fix-pr/SKILL.md:
- Around line 28-51: Update the GraphQL command in the review-thread retrieval
instructions to use gh api graphql --paginate --slurp instead of manual cursor
reruns and accumulation. Add the pagination variable expected by gh, use it in
reviewThreads, and remove the guidance to rerun with -F cursor or manually
combine pages; preserve the existing /tmp/threads.json output and subsequent gh
pr checks command.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: a339ca84-73ab-4193-a205-086071bb36cf

📥 Commits

Reviewing files that changed from the base of the PR and between 72ee6a1 and f9fb940.

📒 Files selected for processing (6)

• .claude/skills/fix-pr/SKILL.md
• .claude/skills/git-commit/SKILL.md
• .claude/skills/github-pr/SKILL.md
• golden/runner.py
• tests/golden/conftest.py
• tests/golden/test_runner.py