Your first pull request

A checklist, not a restatement. Every line links to the section of AGENTS.md that is the rule — the point is that the rule has one home and this page is only a way into it. If the two ever disagree, AGENTS.md is the one that is right (§1.7: fix one of them, never invent a third name).

The commands and their costs are in CONTRIBUTING.md and docs/guides/build-time.md.

#	before you commit	the rule
1	Every degradation you added is LOUD, COUNTED, or it is a defect. A missing file stops the run; a wrong shape asserts; a skipped kernel prints a counter. “The fallback computes the right answer” is the reason this class is expensive here — it means a silent bug reads as a year of lost throughput	§1.1 — the three marks are the whole taxonomy
2	Your mechanism has a gate: an assertion, a counter, or a seam a test can read. If it only exists in the log, it does not exist	§1.1 — “a fused/accelerated arm must be able to show it ran”
3	Every claim you wrote names its evidence: a config, a date, a commit, a file:line. “Verified” and “bit-for-bit” are reserved for a named external source; where none exists, say no external reference exists	§1.4
4	The new mechanism beats its own removal on held-out BPB at a fixed step budget, or it is deleted. A tie deletes the mechanism; 3 seeds per arm. If it is not A/B’d yet, say that on the page, in those words	§1.2 · queue: docs/protocols/AB-PROTOCOL.md
5	No host-device sync crept in: no try_into_scalar, no into_data, no blocking_read, no host-side branch on a device value. Anything a host must know is produced by a device counter and read at a declared cadence. Count on the host; never build a numeric indicator from a bool tensor on device	§1.3
6	You checked what this box cannot do before you relied on it. Precision, 4-D slicing, the allocator high-water, mixed-dtype ops and the shard corpus each have a recorded failure and a recorded reason	§2.1–§2.3, §2.5, §2.6
7	Nothing heavy ran twice at once. One GPU process, one build, under the build lock and the MemoryMax cgroup; free -g avail ≥ 25 before a build or a run	§1.5 · §2.4
8	Your branch is a worktree off a compiling commit (tools/wt.sh), you did not touch a file another agent is in, and you staged only the paths you changed — no git add .	§1.6
9	A document moved: the old path has no stub left, every reference was rewritten in the same commit, python3 tools/check_doc_refs.py is green, and a new document has a line in docs-site/tools/manifest.mjs	§1.8
10	You did not push, force-push, or run a destructive git operation. That needs the owner’s OK; local reversible actions are free	§1.6

The two gates that are easy to forget

• cargo check is not enough for a kernel, a dtype or a device path. The class of bug that only appears on the CUDA backend cannot be caught by a CPU test suite: run cargo test -p backend-parity --features cuda --test backend_parity (§2.1 — it turns “the cast is broken” into a command instead of a paragraph).
• A CPU test can pass while asserting nothing. The measured example is in §3.3: a bit-exactness gate that was RED 976/1000 and whose 1000-case test was feature-gated off the default cell. Read what your assertion can fail before you trust it.

What “done” means here

Full tests green and typecheck/lint clean. “Smoke” is not done. A failing test is a bug in the code, not a test to skip. Nothing is deleted to make a gate green.

Fixing something unrelated in the same diff is a review finding, not a courtesy: report it with file:line and leave it.