feat(drift): cgis_drift has no --exclude/--scope, so the hygiene gate measures test volume and logging density

[zaebee]

cgis_metrics gained --exclude in #234 and --scope in #239. cgis_drift got neither, and on a repo with tests beside the code that turns the hygiene gate into a measurement of something other than architecture.

Measured on Ownima owner-api (2026-10-02)

Ingest root ownima-backend, 19 303 nodes / 96 805 edges, graph healthy (unresolved_ratio 0.1297 against a 0.3 threshold). Ontology: the repo's own docs/ontology/patterns.yaml, 17 domains.

Result: any_critical: true, 13 of 17 domains gate_failed. But only 4 exceed their own drift_tolerance:

domain	drift	tolerance	over?
scripts	0.229	0.11	yes, 2×
common	0.556	0.53	yes
core	0.424	0.39	yes
mcp	0.333	0.32	yes

The other nine fail only on hygiene unresolved_ratio > 0.2. Here is what those unresolved edges actually are, from cgis_validate's top_unresolved on the same graph:

response.json       497
self.client.post    493
self.client.get     349
logger.info         279
logger.exception    212
logger.warning      181
resp.json           143
r.json              113

Two populations, neither architectural:

1. logger.* — logger = structlog.get_logger(__name__), a bound logger. Unresolvable by construction, and a domain's share of these is a function of how much it logs.
2. self.client.post / response.json — the test HTTP client. These exist only because tests are in the graph.

So the per-domain unresolved_ratio ranks domains by test volume and logging density. The node counts show the contamination directly: reservation 1130 nodes, finance 599, vehicle 485, notifications 476, chat 463 — these are mostly test files. app.vocabulary reports unresolved_ratio 0.50 on 39 nodes; app.worker 0.40.

Second symptom: the fit band goes to pure_utility for layered domains

Six domains declared layered_dag or dispatcher come back with pure_utility as fit.nearest_template:

domain	expected	nearest	residual	band
reservation	layered_dag	pure_utility	0.351	weak
vehicle	layered_dag	pure_utility	0.470	none
chat	layered_dag	pure_utility	0.280	weak
finance	layered_dag	pure_utility	0.466	none
notifications	dispatcher	pure_utility	0.284	weak
crud	pure_utility	layered_dag	0.470	none

Test modules are flat and leafy — each test function calls into the app and is called by nobody — so they pull the motif distribution toward pure_utility. When six clearly-layered service domains all land nearest to "pure utility", the basis is describing the test suite, not the design.

Ask

exclude: list[str] and scope: list[str] on cgis_drift, with the same dot-segment / dot-prefix semantics as cgis_metrics so one ontology can be scored the same way twice.

Two notes on semantics, since drift differs from metrics here:

• Exclusion must happen before the motif census and before unresolved_ratio, not as a row filter on the output — both are aggregates over the domain's subgraph, so filtering afterwards changes nothing.
• Node counts should be reported post-exclusion in actual, so a reader can see what was scored. Right now node_count: 1130 for reservation reads as a domain of 1130 production symbols.

Why it matters beyond cosmetics

A gate whose failures are dominated by an artifact teaches people to ignore it. We cannot put cgis_drift in CI in this state: 13 red domains, 9 of them red for reasons no code change would fix. With exclusion we would gate on the 4 that are real.

Related: #234 (exclude in metrics), #239 (scope in metrics), #178 (a mis-targeted ontology reading clean — the inverse failure), #170 (status semantics / hygiene gate).

[zaebee]

Correction that makes this cheaper than I filed it: the data to exclude tests is already in the graph and the scorer ignores it. nodes carries an is_test column, it is populated, and it is accurate.

Measured on the same graph:

SELECT is_test, COUNT(*) FROM nodes WHERE id LIKE 'app.domains.reservation%' GROUP BY is_test;
-- 0 |  201
-- 1 |  929
SELECT is_test, COUNT(*) FROM nodes GROUP BY is_test;
-- 0 |  8208
-- 1 | 11095

So app.domains.reservation's reported node_count: 1130 is 201 production nodes and 929 test nodes — 82% tests. Whole-graph it is 57%.

That reframes the ask. This is not "add an exclude parameter so a caller can filter"; it is "the scorer aggregates over a column it already populates and does not read". Two consequences:

1. The fix can be a default rather than a flag: cgis_drift scoring is_test = 0 unless asked otherwise would be the honest default, since no ontology describes a test suite's architecture. An exclude/scope pair is still worth having for vendor and generated code, but it is no longer the thing standing between this tool and CI.
2. It explains the magnitude. A hygiene gate at 0.2 fails almost everything because the denominator is four-fifths test code, and tests are where self.client.post and response.json live — the two largest unresolved populations in the whole graph.

is_generated is beside it in the same table and presumably as usable, which would take app.gen.* out too.