compass(docs): pin 16's first CPU-tier census and 08's Revised paragraph to their commits

[jgong5]

Closes #344

Dev record

What was done. I chose to pin, not delete. Each figure in P0.1's CPU-tier bullets in 16, and in the "Revised 2026-09-20" paragraph in 08, now names the commit it was measured at. Figures I could not tie to a commit are dropped. There is one commit, f8b2b2deb, on d96da1086. It changes 27 lines and removes 29, docs only.

16_execution_plan.md

Sentence (after)	Evidence for the commit
"130 of the 189 test files at 105ca4197, 4030 passed, 0 failed, 149 skipped, 3 xfailed, rc=0"	PR #6's dev record gate table: its branch column is 105ca4197 with 4030 / 0 / 149 / 3 and GATE_CPU_RC=0. #6's reviewer reproduced the same counts at 9664ed025. git ls-tree at 105ca4197 gives 189 test files: 30 in tests/plugin/, 29 excluded, 130 in the tier.
"105ca4197 is P0.1's head and landed as 4c16792d9, the same tree"	git rev-parse 105ca4197^{tree} 4c16792d9^{tree} gives fbcb03674 twice.
"Six runs at … d737f15e7 and 7ff80cc4b, gave the same counts, and the clock read 25.6-30.8 s of pytest inside 31.6-36.9 s of wall"	Commit 8c0ee374f's message and diff: "six runs, three at d737f15 and three at 7ff80cc … the same 4030 / 0 / 149 / 3 in all six". Both commits are pre-restack commits of #6's branch (git merge-base --is-ancestor fails for both; both resolve on jgong5/ATOM). The old 25.4-31.7 / 31.1-37.8 range came from "seventeen runs across ten commits" (#6's body), and no record names those commits, so it is replaced rather than pinned.
"The 4030 at 105ca4197 is 3956 ATOM + 74 tests/compass/"; "The 3956 is the same gate at the base, 042aad97d"	Same table, control column: 042aad97d, 3956 / 0 / 149 / 3. #6's reviewer measured tests/compass at 66 + 8 = 74 passed.
"4005 was the same gate at 3afcb4880, with tests/compass at 49 tests"; "+17 … the other 8"	4005 at 3afcb4880 is unchanged text (commit b78e0bc97's message: "4005 … at 3afcb48 on node 18"). The arithmetic is 49 + 17 + 8 = 74. 3925 and 4022 are dropped. git log -S traces them to b78e0bc97 and 5ca21de2b, and neither commit names a tree for them. They survive only in scripts/compass/README.md, which is outside this file set.
"Of the 130 files handed at 105ca4197, 22 collect no test at all"	gpu_gate_triggers.txt's header at 4c16792d9 reads "Measured at fada742 … CPU tier 108 files collect >=1 test, 22 collect none". fada7424e and 105ca4197 have the same tests/ (105ca4197 only restamps that header).
"measured per file by P0.2 at b963c9411, where the same 22 sat inside a 128-file gate whose 149 skips were 68 distinct reasons … Both carry to 105ca4197: its tests/ and code differ from b963c9411's only by the two tests/compass/ files above, which collect tests and skip none"	4c16792d9^:16_execution_plan.md states the P0.2 bullet at b963c9411: 128 files handed, 22 collect none, 149 skips = 68 reasons. git diff --stat b963c9411 105ca4197 -- tests atom shows only the two tests/compass/ files plus atom/compass docs. There are 149 skips on both sides of #6's table.
"header included (at fada7424e: 189 test files, 130 CPU tier, …)"	The header's own "Measured at fada742…" line.

08_validation_protocol.md

Sentence (after)	Evidence
"At fada7424e there are 189 test files under tests/, and the 29 of the 159 outside tests/plugin/ that scripts/compass/cpu_gate_exclude.txt lists reach the driver, each run alone: 28 at collection … plus 1 … hipHostMalloc failed: 100"	git ls-tree at fada7424e: 189 / 30 / 29, so 159 are non-plugin. I re-measured the driver claim on node 18, in xiaobizh_n18_cpu, on a git archive of fada7424e with atom.__file__ = /tmp/i344gates/fada/ATOM/atom/__init__.py. I ran one pytest <file> per process under timeout -k 10 150. All 28 GENERATED entries give rc=2, ERROR collecting and RuntimeError: Get GPU arch from rocminfo failed. test_lmcache_offload_disk_integration.py collects, gives 2 failed in 5.69s with RuntimeError: hipHostMalloc failed: 100 twice, and then does not exit, so the timeout killed it (rc=124). I wrote "that cpu_gate_exclude.txt lists" rather than a bare "29 reach the driver" because two tier files also reach rocminfo when run alone. The row at :49 already says so.
"collecting it alone at 236abfd9a, and again at fada7424e, in xiaobizh_n18_cpu gives 153 tests collected, 7 errors in 1.45s, rc=2"	Same session: pytest tests/plugin --collect-only gives rc=2 and 153 tests collected, 7 errors in 1.45s. The errors are 3 rocminfo (the same three files), 1 No module named 'sglang' and 3 cannot import name 'fused_gdn_gating'. That makes "at least 32 of 189" a reading at fada7424e. tests/plugin is the same tree object, fb7ee190d, at both commits.
"tests/plugin/'s 30 files at fada7424e"	git ls-tree.
:49 "Green, GATE_CPU_RC=0, read at bdd244c57 in xiaobizh_n18_cpu" (r4087085010)	#345's gate table: control bdd244c57, GATE_CPU_RC=0.
:52–54, only the test_dp_* / test_forward_mode sentence kept (r4087085189, shrink:)	Net −2 lines.

What was decided that the brief did not cover.

• Dropping 3925 and 4022, not pinning them. No record names their trees. Keeping them would leave two pass counts with no commit.
• Pinning the header counts at 16:195. It was a test-file count with no commit. It was a one-phrase fix.

What surprised me. scripts/compass/README.md:113 says the 4030 row's "commit not recorded". It was recorded: #6's own gate table measured it at 105ca4197. That README row, and its 3925 / 4022 paragraph, are outside this file set, so they are left undone here. They need a follow-up issue.

Also left undone. 08:66 ("gives 4 passed in the GPU container") is a pass count with no commit. It is outside the issue's regions, and it needs a GPU-container reading to pin.

Named result

• grep -rn -E "4030|130 of" atom/compass/design/ at f8b2b2deb finds 3 lines, all in 16_execution_plan.md (117, 124, 127), and each carries 105ca4197. 130 of 189 gives 0 hits, and 3925 / 4022 give 0 hits.
• No sentence in the changed regions states a test-file count, pass count or timing without naming its commit (tables above).

Gate 1: ATOM's suite, unmodified, as a delta against a control

Node 18, xiaobizh_n18_cpu, one run at a time. Each tree was staged with git archive into /tmp/i344gates/{control,branch}/ATOM via docker exec -i … tar -x. A content md5 of every staged file matched on both ends. atom.__file__ was asserted under each root before its run. Each run used the tree's own scripts/compass/gate_cpu.sh --junitxml=… under timeout -k 10 1800, not piped. .compass-changed holds this PR's 2 files on both sides.

	tree	commit: printed	result	GATE_CPU_RC
control	tip f21492580 (the tip moved past d96da1086 to #329 while I worked)	f21492580 (stamp)	5258 passed, 155 skipped, 3 xfailed, 185.74 s	0
branch	merged 6ee3a6c07, stamp 43549daee (git commit-tree, parents tip + head)	43549daee (stamp)	5258 passed, 155 skipped, 3 xfailed, 188.68 s	0

• Node-id delta: 0. Both junits hold 5416 ids. The diff is empty on ids alone and on id + outcome.
• The timing classes passed on both sides, so none needed a re-run: TestTheRegionIsNotCopiedPerChunk, TestNoSizeAtWhichACallStopsBeingOne::…[minimax] and test_gc_utils.py::test_freezing_twice_is_additive_and_harmless.
• Production lines: 0. Test lines: 0. Doc lines: +27 / −29.
• git merge-tree --write-tree f21492580 f8b2b2deb = 6ee3a6c07354b990444184a4edabd41b49d42cbf, clean. compass(runner): the speculative refusal reads its field with no default; refuse gpu_memory_utilization outside dict keys #329 touches only atom/compass/runner/overrides.py and three tests/compass/ files.

🤖 Generated with Claude Code

[jgong5]

Non-blocking. Principle 8 ("Every claim carries its measurement. A number without a source is a defect").

This sentence attaches one wall time, 1.45s, to two runs: one at 236abfd9a and one at fada7424e. I ran the same collection on node 18, in xiaobizh_n18_cpu, on a git archive of fada7424e, with atom.__file__ = /tmp/pr349r1-gates/fada/ATOM/atom/__init__.py. python -m pytest -q --collect-only tests/plugin gave rc=2 and 153 tests collected, 7 errors in 1.41s. The 7 errors split the same way the text says: 3 rocminfo, 1 No module named 'sglang' and 3 cannot import name 'fused_gdn_gating'.

The counts and the rc reproduce. The time is noise, so it cannot be a reading of both commits. Either drop in 1.45s, or say it is the 236abfd9a reading. For example: "gives 153 tests collected, 7 errors, rc=2 (1.45 s at 236abfd9a)".

[jgong5]

Non-blocking. Principle 7 ("Never report an aggregate without its decomposition").

This PR narrows the numerator to "the 29 … that cpu_gate_exclude.txt lists". The PR body gives the reason: "two tier files also reach rocminfo when run alone". I measured both at fada7424e on node 18, in xiaobizh_n18_cpu, on a git archive with atom.__file__ under /tmp/pr349r1-gates/fada/ATOM. Each was run alone:

• tests/test_postprocess_width.py: rc=2, 1 error in 0.76s, Get GPU arch from rocminfo failed.
• tests/test_v4_checkpoint_slot_copy.py: rc=2, 1 error in 0.94s, the same rocminfo error.

So the per-file figure measured at fada7424e is 29 + 2 + 3 = 34 of 189. "At least 32" is still true. But this sentence explains the floor by pointing at the 23 unrun plugin files alone, and leaves out a part that has been measured. Suggestion: "29 + 3 = 32, plus the 2 tier files the table below names, so at least 34; the other 23 plugin files …". Or keep 32 and name the 2 as a known excess.

[jgong5]

Non-blocking. Principle 8, and ponytail shrink:.

The citation is correct. #6's gate table has 042aad97d in its control column: 3956 passed, 0 failed, 149 skipped, 3 xfailed, rc=0. That control was not the base tree as it stands, though. #6's body says: "Control is fork/feature/atomcompass_new at 042aad97d with only scripts/compass/ copied in". git ls-tree 042aad97d scripts/ has no scripts/compass/, so a reader who checks out 042aad97d finds no gate to run. The words "the same gate" are only literally true.

shrink: L126-129. This sentence and the one that ends at L129 ("the P0.2 readings above are the same suite without them") make one point twice: the 3956 half has no tests/compass/. One sentence can carry the pin and the control's construction. For example: "The 3956 is #6's control, 042aad97d with #6's scripts/compass/ copied in and no tests/compass/, as in the P0.2 readings above; 105ca4197 adds tests/compass/test_cpu_gate_exclude.py and tests/compass/test_gate_gpu_surplus.py, so the file count moves 128 → 130." That is -1 line.

[jgong5]

Review cycle 1: APPROVE at head f8b2b2deb98d1e38ed4e422bbf681861b101068d.

No finding is blocking. There are 3 inline comments, all non-blocking: r4087445424 (08:34), r4087445580 (08:39) and r4087445720 (16:126). There is 1 further non-blocking finding with no line. It is below, as finding N1.

Before reviewing, I read the eight design principles in atom/compass/design/README.md and atom/compass/AI_DEV_RULES.md, both at tip f21492580. This PR's content is provenance, so the review is principle 8, "Every claim carries its measurement. A number without a source is a defect". I checked every citation against git, against #6's thread, and against node 18.

Citation by citation

#	Claim	Commit	Evidence I checked	Verdict
1a	105ca4197 is #6's head	105ca4197	GET pulls/6: head.sha=105ca41975d9…, merged=true, merge_commit_sha=4c16792d9…	true
1b	It is the same tree as landed 4c16792d9		rev-parse ^{tree}: fbcb03674 twice	true
1c	#6's gate table gives 4030 / 0 / 149 / 3 at that commit		#6's body, "Gates" table, branch column 105ca4197: 4030 passed, 0 failed, 149 skipped, 3 xfailed, GATE_CPU_RC 0	true
1d	189 test files = 30 plugin + 29 excluded + 130 tier		git ls-tree at 105ca4197: 189 / 30 / 29 non-comment lines in cpu_gate_exclude.txt / 130	true
2a	3956 at 042aad97d	042aad97d	#6's control column: 3956 / 0 / 149 / 3, rc 0. That control is 042aad97d with #6's scripts/compass/ copied in, because 042aad97d has no scripts/compass/ (r4087445720)	true, with that caveat
2b	4005 at 3afcb4880	3afcb4880	b78e0bc97's message: "4005 … 3956 ATOM + 49 tests/compass … at 3afcb48 on node 18"	true
2c	49 + 17 + 8 = 74		4005 − 3956 = 49. f3f584ea4: triggers 13 → 30, so 49 → 66, "4005 -> 4022 (3956 + 66)". #6's body: "test_gate_gpu_surplus.py adds 8". 66 + 8 = 74 = 4030 − 3956	true
3	Six runs at d737f15e7 and 7ff80cc4b, 25.6-30.8 s of pytest inside 31.6-36.9 s of wall	8c0ee374f	The message says "six runs, three at d737f15 and three at 7ff80cc … the same 4030 / 0 / 149 / 3 in all six". The diff gives the same ranges. Both commits come before the restack (neither is an ancestor of 105ca4197, and both resolve on the fork). 8c0ee374f is itself in #6's landed ancestry	true
4	3925 and 4022 cannot be attributed; "neither commit names a tree"	b78e0bc97, 5ca21de2b	-S finds the commits that removed them, not the ones that introduced them. Both figures can be attributed. See N1	false (non-blocking)
5a	"22 collect no test" at 105ca4197	fada7424e	The gpu_gate_triggers.txt header at 105ca4197 / 4c16792d9 reads "Measured at fada742… CPU tier 108 files collect >=1 test, 22 collect none". fada7424e:tests = 105ca4197:tests = 6ca32ac54, and the only diff between them is that header line	true
5b	149 skips = 68 reasons at b963c9411, carried to 105ca4197	b963c9411	042aad97d:16_execution_plan.md states it at b963c9411. git diff --stat b963c9411 105ca4197 -- tests atom gives only tests/compass/test_cpu_gate_exclude.py, tests/compass/test_gate_gpu_surplus.py and atom/compass/ docs. Outside those, only CLAUDE.md and scripts/compass/ differ, and there is no conftest or pytest config change. The two added files skip none: 149 skips on both sides of #6's table. 83daf636d, 042aad97d and 105ca4197 also have identical ATOM tests	true
6	16:195 header counts are at fada7424e	fada7424e	Header lines 40-51: 189 = 30 + 29 + 130, 108 collect, 47 candidates, 99 covered	true
7	08's Revised paragraph at fada7424e: 189 / 159 / 29 / 30, and the listed 29 reach the driver	fada7424e	ls-tree: 189 / 30 / 29. Node 18 spot-check below	true. The 1.45s timing is not (r4087445424)
8	08:49 read at bdd244c57	bdd244c57	#345's gate table: control bdd244c57, GATE_CPU_RC 0	true
9	08:52-54 shrunk, net −2		The diff removes 3 lines and adds 1	true
10	The grep finds 3 lines, all at 105ca4197		Reproduced below	true
11	Gate 5258 = 5258, node delta 0		Reproduced below	true
12	#350 filed		Issue #344's last comment names #350	true

N1 (non-blocking, no line). 3925 and 4022 can be attributed, and the 3925 decomposition that was removed was wrong

This falls under principle 8. Dropping both figures is allowed by the brief ("pin … or" drop), so the tree is right. The PR body's reason for dropping them is wrong, though, and #350's scope repeats that reason.

• 3925 is at 83daf636d. 9c8df1328 (P0.2) says "Measured at 83daf63 on node 18 … a 125-file CPU gate that is genuinely green -- 3925 passed". 9120a96a3's gate_cpu.sh prints "Baseline at 83daf63 is 3925 passed". At 83daf636d there are 187 files and 30 plugin files, so 157 − 32 exclusions = 125, with no tests/compass/. git ls-tree 83daf636d -- tests/compass is empty. Its ATOM tests are identical to 105ca4197's. So the sentence this PR removed, "3925 was 32 exclusions with tests/compass at 35 tests", was false. The 35 belonged to eb1d6d4fa's 3960 ("3960 = the 3925 baseline + exactly the 35 tests this task adds"). The real decomposition is 3925 + 31 = 3956. The 31 are the three re-measured files, 16:108's "31 passed in 0.18 s".
• 4022 is at 71d2a1ac2. f3f584ea4's message reads "CPU tier 4005 -> 4022 (3956 + 66)". The trigger header that 5ca21de2b removed begins "Measured at 71d2a1a… over this tree: 188 files … 129 CPU tier … A run of the same set reports … 4022 passed + 128 skipped + 3 xfailed". At 71d2a1ac2, ls-tree gives 188 / 30 / 29 exclusions, with 1 tests/compass file and 30 triggers, so 3956 + 66 = 4022.

Action: no commit is needed here. Please post a correction on #350: scripts/compass/README.md can pin 3925 at 83daf636d and 4022 at 71d2a1ac2, rather than drop them. The wrong "tests/compass at 35" decomposition lived only in 16:129-130 at the tip, which this PR removes. scripts/compass/README.md:125 at the tip says only "3925 at 32 exclusions", which is correct, so it needs the pin and nothing more.

Driver spot-check at fada7424e (node 18, xiaobizh_n18_cpu)

I staged a git archive of fada7424e into /tmp/pr349r1-gates/fada/ATOM, with matching tar md5 at both ends. atom.__file__ = /tmp/pr349r1-gates/fada/ATOM/atom/__init__.py, which I asserted before any run. Each run was one pytest per process under timeout -k 10 150.

file (run alone)	rc	result
tests/test_dcp_topk.py (GENERATED)	2	1 error in 0.63s, Get GPU arch from rocminfo failed
tests/test_eplb_module_c.py (GENERATED)	2	1 error in 0.69s, rocminfo
tests/model_ops/test_balance_router_logits.py (GENERATED)	2	1 error in 0.68s, rocminfo
tests/plugin --collect-only	2	153 tests collected, 7 errors in 1.41s: 3 rocminfo, 1 No module named 'sglang', 3 cannot import name 'fused_gdn_gating'
tests/test_postprocess_width.py (tier)	2	1 error in 0.76s, rocminfo
tests/test_v4_checkpoint_slot_copy.py (tier)	2	1 error in 0.94s, rocminfo

The wording "the 29 … that cpu_gate_exclude.txt lists reach the driver" is exactly true, and it is right to qualify it. The last two rows show that 2 more files reach it too. That affects the "at least 32" explanation at 08:39 (r4087445580). The plugin time was 1.41 s, not 1.45 s (r4087445424).

Grep and sweep

• git grep -n -E "4030|130 of" <rev> -- atom/compass/design/, at both f8b2b2deb and the merged tree 6ee3a6c07, finds 3 lines: 16:117, 16:124 and 16:127. Each one names 105ca4197. 130 of 189, 3925 and 4022 give 0 hits.
• I swept both files at the head for N passed|failed|skipped|files|tests and for N s timings:
  • In the changed regions, every figure names a commit. The one exception is the 1.45s shared by two commits (r4087445424).
  • The 16:115 "(30 files) … the 29" lead-in is pinned by the same sentence's 105ca4197, and it also holds at the tip.
  • 16:193's "30 paths" holds at 105ca4197 and at the tip.
  • Outside the changed regions, the only unpinned count is 08:65, "gives 4 passed in the GPU container", which is already filed as compass(docs): scripts/compass/README.md still says the 4030 row has no commit, and lists 3925 / 4022 unpinned #350.
  • 08:64 and 08:68 ("no tests ran", "1 skipped") are pinned at 83daf636d by 08:76-77.

ponytail-review

16_execution_plan.md L126-129: shrink: "The 3956 is the same gate at the base…" and "the P0.2 readings above are the same suite without them" say "3956 has no tests/compass" twice. One sentence, and it names the control's copied scripts/compass (r4087445720).

net: -1 lines possible.

Gate: merged tree, one run

• Tip read again just before posting: f21492580525601f8e6e76efcd4e89b9e95eda1a. The PR head is f8b2b2deb, still a draft, with no labels.
• git merge-tree --write-tree f21492580 f8b2b2deb = 6ee3a6c07354b990444184a4edabd41b49d42cbf. That is clean, and the same as the developer's.
• Stamp: git commit-tree 6ee3a6c07 -p f21492580 -p f8b2b2deb = 824151dce. No ref was created.
  • .compass-commit = 824151dce….
  • .compass-changed holds the 2 doc files.
• Staged with git archive + docker exec -i … tar -x into /tmp/pr349r1-gates/merged/ATOM, with matching tar md5 at both ends. atom.__file__ = /tmp/pr349r1-gates/merged/ATOM/atom/__init__.py. The tree's own scripts/compass/gate_cpu.sh --junitxml=… ran under timeout -k 10 1800, not piped, with load average about 31.
• Printed commit: 824151dce (stamp) and gpu: not required (.compass-changed stamp).
• 5258 passed, 155 skipped, 3 xfailed, 189.98 s; GATE_CPU_RC=0. The junit has 5416 test cases, 0 failures and 0 errors. This equals the developer's control (f21492580) and branch counts exactly.
• No timing class failed, so none needed a re-run: TestTheRegionIsNotCopiedPerChunk, TestNoSizeAtWhichACallStopsBeingOne::…[minimax] and test_gc_utils.py::test_freezing_twice_is_additive_and_harmless.

Next task in this area should watch

#350 should work from N1's provenance rather than from this PR body's "unattributable". That means pinning 3925 at 83daf636d and 4022 at 71d2a1ac2 in scripts/compass/README.md:125-126, not dropping them.