feat(tests): stamp testtype__/model__ JUnit tags so tier coverage is recorded

[ashokponkumar]

Problem

hf-adapters emits no JUnit result tags at all — record_property appears in no file in the repo. So the CI/CD warehouse has zero tier-coverage data for it: 0 rows carrying a testtype__ tag, against 3,311 for spyre-inference. Anything asking "has this artifact already been tested at tier X?" can only answer "no" here, so no hf run can ever be reused.

Change

Each test case now carries testtype__<tier> for every tier its suite belongs to, plus model__<id> from the parametrization. Emitted as <property name="tag" value="namespace__value"/> — the shape the ingest already on main parses, with no ingest change needed (verified: it writes one hf_run_properties row per tag).

The tier can't be recovered at test time — CI calls the per-suite make targets directly (never make tests), and a suite target doesn't know which tier invoked it. So the suite is declared with a new --suite <key> option, and tests/_tier_tags.py maps it to the tier set.

--suite rather than an env var because that's this repo's existing idiom (--run-slow, --model-path); no conftest here reads env vars for configuration.

Membership is declared, never inferred from a ladder

11 of the gated suites are [regression, trunk, unit] with no integration. Expanding "in unit ⇒ in everything above" would claim integration coverage for suites that never ran under it and silently skip real tests. A test asserts a ladder would still over-claim these suites, so the shortcut can't creep back in.

Worth reviewer attention: the mapping already exists twice, and the copies disagree

Source of truth here is each suite job's if: gate in _test_matrix.yaml — that's what actually decides whether a suite runs. But the Makefile's tests target carries a second, hand-maintained copy, and the two have already drifted (pre-existing, not from this PR):

suite	workflow if:	Makefile case
clip	regression, trunk, unit	absent from every suites= list
multicard_smoke	regression, trunk, unit	no target at all
adapter_coverage	no tier gate (runs always)	regression, trunk, unit
model_components	integration, regression, trunk, unit	regression, trunk

model_components is the interesting one: Makefile:8 documents it as an integration suite and CI agrees — the Makefile's own case block is the stale copy.

A test pins this known divergence, so a new one fails rather than being absorbed silently. I deliberately did not reconcile them: that changes which suites make tests runs, and belongs in its own change.

Five suites are deliberately untagged

Each for a checked reason — tier_tags() returns [], so they still get a model tag:

• perf — scaffold; echoes an empty JUnit with no <testcase> elements
• edge_cases — gated on inputs.edge_cases_only, never a tier
• multicard_smoke — runs a torchrun script, emits no JUnit
• adapter_coverage — runs --noconftest, so the fixture can't bind; passing --suite there is a hard argument error (confirmed)
• model_module — delegates to oot_framework/run_test.sh, which does its own marker-based tagging

Verification

Against the real conftest in a synced venv (not a stub):

• 18 cases each carry the four token_compare tiers
• an embed_compare run carries three (correctly no integration) plus model__…
• omitting --suite emits zero tags rather than failing
• a mutation closing the ladder fails the tests
• 18 new tests; ruff 0.15.6 + black 26.3.1 + the repo's own pre-commit hooks all pass

Scope

Producer only — no change to which tests run, only to what the JUnit XML records. Pairs with torch-spyre/spyre-inference#847, which fixes the same class of gap on that side.