compass(tests): pin the geometry refusal for every field it reads

[jgong5]

Closes #301

What changed

test_a_config_with_no_hidden_size_refuses_naming_the_field becomes test_a_config_missing_a_geometry_field_refuses_naming_that_field, parametrized over the four fields _geometry reads in ModelTerms.declared_for_m1 (hidden_size, intermediate_size, head_dim, max_position_embeddings). Each case deletes its field and asserts MemoryRefusal with .what naming that field. The remedy assertion is unchanged.

File set: tests/compass/test_memory_readings.py only. Production lines: 0. Test lines: +6 / -3 (net +3), inside the +2-5 estimate.

Named result (node 18, xiaobizh_n18_cpu)

Mutations are line-count-preserving swaps of atom/compass/memory/readings.py (366 lines each), run on the tip (b3078336d) and the head (b736fd8ba), tests/compass/test_memory_readings.py:

mutant	edit	tip	head
null	docstring wording on _geometry only	47 passed	50 passed
M2	line 76: `{name}` -> `hidden_size` in the message	47 passed	3 failed, 47 passed
M4	line 137: int(_geometry(config, "intermediate_size")) -> int(getattr(text, "intermediate_size", 0))	47 passed	1 failed, 49 passed

Failing node ids at the head:

• M2: tests/compass/test_memory_readings.py::test_a_config_missing_a_geometry_field_refuses_naming_that_field[intermediate_size], [head_dim], [max_position_embeddings], each AssertionError on .what (states no hidden_size`` against the deleted field). [hidden_size] stays green, as it must.
• M4: ...::test_a_config_missing_a_geometry_field_refuses_naming_that_field[intermediate_size], Failed: DID NOT RAISE MemoryRefusal.

Original readings.py restored after each run and md5-checked (438f6578b550 on both trees).

Gates

Gate 1: ATOM suite, unmodified, CPU tier (scripts/compass/gate_cpu.sh from each tree, git archive snapshots with .compass-commit / .compass-changed stamps, atom.__file__ asserted under the staged root):

tree	atom.file	result	GATE_CPU_RC
control b3078336d	/tmp/i301gates/ctl/ATOM/atom/__init__.py	5206 passed, 155 skipped, 3 xfailed	0
head b736fd8ba	/tmp/i301gates/head/ATOM/atom/__init__.py	5209 passed, 155 skipped, 3 xfailed	0

Delta +3 passed, 0 failed. Node-id delta (collect-only on the changed file, 47 -> 50): removed test_a_config_with_no_hidden_size_refuses_naming_the_field; added test_a_config_missing_a_geometry_field_refuses_naming_that_field[hidden_size|intermediate_size|head_dim|max_position_embeddings]. No flaky-class failures on either side.

git merge-tree --write-tree b3078336d b736fd8ba = 17e76b25c472beca4c0c65e47646fe56944614bd = b736fd8ba^{tree}.

Gate 2: the change is itself a CPU-only test in tests/compass/; it exercises _geometry and ModelTerms.declared_for_m1, which this PR does not touch.

Gate 3: the named result above.

Gate 4: reviewer to be dispatched by the coordinator.

Dev record

• No surprises. The fixture's text_config carries all four fields, so delattr removes a real attribute in every case.
• The pending declared_for_m1 rename noted in the brief has no open PR; nothing to coordinate.
• ruff format and ruff check clean on the changed file.

🤖 Generated with Claude Code

[jgong5]

Review: APPROVE at head b736fd8bafbf05df42f871096f5990d83669417e

Verdict: APPROVE, covering b736fd8bafbf05df42f871096f5990d83669417e (tree 17e76b25c472beca4c0c65e47646fe56944614bd). There are no blocking findings and no inline comments.

I read the eight design principles and AI_DEV_RULES.md first. A previous reviewer stopped on an API error before posting, so this is the first review on this thread.

1. Field completeness (principle 6)

At the tip b3078336d, atom/compass/memory/readings.py calls _geometry exactly four times, at lines 136-139: hidden_size, intermediate_size, head_dim and max_position_embeddings. The parametrize list is exactly that set.

The module reads two other config fields, and both are already pinned by their own tests:

• dtype goes through _dtype and refuses. test_a_config_with_no_dtype_refuses_rather_than_assuming_one pins it.
• partial_rotary_factor defaults to 1.0 on purpose (line 147), and the default is labelled in the term source. test_an_absent_partial_rotary_factor_says_so_in_the_table pins it.

getattr(config, "text_config", config) is a structural lookup, not a value default. Nothing is missing.

2. Named result, reproduced (principle 8)

Measured on node 18, xiaobizh_n18_cpu, file tests/compass/test_memory_readings.py.

• Trees are git archive snapshots staged at /tmp/rv303b/{tip,head}/ATOM, and atom.__file__ resolves under each staged root.
• Each mutant is a one-line swap of readings.py: 366 lines before and after, and diff shows one changed line.
• The original file was restored after each run and md5-checked (438f6578b550).

mutant	edit	tip b3078336d	head b736fd8ba
none		47 passed	50 passed
null	line 71, docstring wording only	47 passed	50 passed
M2	line 76, `{name}` → `hidden_size`	47 passed	3 failed, 47 passed
M4	line 137, int(getattr(text, "intermediate_size", 0))	47 passed	1 failed, 49 passed
M5 (reviewer)	line 138, int(getattr(text, "head_dim", 128))	47 passed	1 failed, 49 passed
M6 (reviewer)	line 139, int(getattr(text, "max_position_embeddings", 40960))	47 passed	1 failed, 49 passed
M7 (reviewer)	line 74, if value is None and name == "hidden_size":, so the refusal fires only for the field the old test deleted	47 passed	3 failed, 47 passed

Failing node ids at the head, all tests/compass/test_memory_readings.py::test_a_config_missing_a_geometry_field_refuses_naming_that_field[...]:

• M2: [intermediate_size], [head_dim], [max_position_embeddings]. Each is an AssertionError on .what, reading states no hidden_size`` against the deleted field. [hidden_size] stays green, as it must.
• M4: [intermediate_size], with Failed: DID NOT RAISE <class 'atom.compass.memory.readings.MemoryRefusal'>.
• M5: [head_dim], with DID NOT RAISE MemoryRefusal.
• M6: [max_position_embeddings], with DID NOT RAISE MemoryRefusal.
• M7: [intermediate_size], [head_dim], [max_position_embeddings], each TypeError: int() argument must be ... not 'NoneType'.

Every mutant is green on the tip, so the old single-field test saw none of them. On the head, each field's case fires on its own. The pin is not inert.

3. Tree that will land

• git merge-tree --write-tree b3078336d b736fd8ba gives 17e76b25c472beca4c0c65e47646fe56944614bd. That equals b736fd8ba^{tree}, so the head is already based on the tip.
• For the stamp I built a commit object with git commit-tree 17e76b25c -p b3078336d -p b736fd8ba, which gives 1ffa6fc39fb1c0124fe7dd86aa0ef655e885a7ca.
• .compass-commit and .compass-changed were written from that commit. The changed set is tests/compass/test_memory_readings.py.

The gate was the tree's own scripts/compass/gate_cpu.sh, run once on node 18 from /tmp/rv303b/merged/ATOM:

atom:   /tmp/rv303b/merged/ATOM/atom/__init__.py
commit: 1ffa6fc39 (stamp)
gpu:    not required (.compass-changed stamp)
5209 passed, 155 skipped, 3 xfailed, 16 warnings in 189.07s
GATE_CPU_RC=0 PASSED

That is 5209 passed, identical to the dev record's head figure. It is +3 over the dev record's control of 5206 at b3078336d: one test was removed and four parametrized cases added. Nothing failed, including the known timing-flaky classes in tests/entrypoints/test_stream_marker_properties.py. ruff check and ruff format --check (0.16.7) are clean on the changed file.

4. No design-doc references

I scanned tests/compass/test_memory_readings.py in full at the head for D<n>, P<n>.<n>, T<n>, W<n>, "principle", "Gate N", doc numbers and design/. There were no hits.

5. ponytail-review (principle 3)

The diff is +6/−3 and replaces a single-field test with a four-value parametrize and an f-string. Nothing is dead or speculative, and no shorter form of the same check exists.

Lean already. Ship.

Notes (non-blocking)

• Hand-kept field list (principle 3). A fifth _geometry call added later would stay uncovered until someone extends this list. Deriving the list from the source would be machinery for a four-item constant, so I accept the list as written. The next task that adds a geometry field should extend this parametrize list in the same PR.
• Dev-record wording (principle 8). The record says "the fixture's text_config carries all four fields". The qwen fixture is PretrainedConfig.from_dict(raw["text_config"]), a flat config with no text_config attribute, so _geometry reads the config itself. The substance holds, as M4-M6 show: delattr removes the attribute _geometry actually reads.

No blocking issues.