test: reconcile ent#89 tests with ent#128's compatibility contract

[trinity-ability]

What

Four unit tests are failing on dev right now. This fixes them.

They are not a defect in either PR that produced them: #1899 (ent#128) and #1946 (ent#89) were each green in isolation, and the failures only exist in their combination. Nothing in the production code is wrong — the four assertions describe a contract that #1899 deliberately changed, and #1946 was written before that landed.

Why it reached dev

dev's required checks are Analyze (python), Analyze (javascript-typescript), schema-parity and verify-non-root. None of them runs a unit test. The regression diff job does, but it reports after the required set goes green — so auto-merge had already landed #1946 by the time the failures were visible.

regression diff on the merge commit reported exactly these four, against a base with zero failures:

## ❌ New failures introduced by HEAD (4)
- [F] test_compatibility_checks.TestReportDirectionOnAFailingValidator::test_persisted_check_error_does_not_replay_as_clean
- [F] test_compatibility_checks.TestRunStaticLogsItsSwallow::test_a_raising_check_is_logged
- [F] test_ent89_template_schedules.TestCatalogSurface::test_a_malformed_block_does_not_raise_out_of_either_builder
- [F] test_ent89_template_schedules.TestCatalogSurface::test_local_builder_surfaces_normalized_schedules

All six pytest jobs (base and head × 3 seeds) themselves passed — they emit JUnit XML and let regression diff be the gate.

The four

1. test_a_raising_check_is_logged

run_static's per-check swallow used to record _skip(..., "check_error"). #1899 changed it to _fail(..., {"check_error": ...}) so that a crashed check is actually counted by _counts — without that, a raise inside a HARD check dropped hard_count and flipped overall_status to compatible on a broken agent.

The test still read detail["skip_reason"], which no longer exists. It now asserts the FAIL shape, which is the fix — only a counted status reaches hard_count/soft_count.

Worth noting the file was already internally inconsistent: a sibling test in the same class asserts t018["status"] == "fail" and was passing.

2. test_persisted_check_error_does_not_replay_as_clean

This asserted _counts(...)["soft_count"] == 0 for a persisted skipped + check_error row — i.e. it pinned the bug as expected behaviour. #1899's _did_not_pass counts exactly that row as a finding, on purpose.

Expectation flipped to 1, and the docstring rewritten so the test keeps real teeth: a row persisted by an older build, before the swallow started returning fail, still must not replay as a clean bill of health on a stopped-agent read.

3 & 4. _build_local_template(...)

#1899 gave it a required keyword-only is_bundled parameter (it selects the credential-metadata trust label, so an operator-uploaded template does not inherit the bundled catalog's label). Both ent#89 call sites now pass it. Orthogonal to the schedules: block under test.

Verification

$ pytest tests/unit/test_compatibility_checks.py \
         tests/unit/test_ent89_template_schedules.py \
         tests/unit/test_ent128b1_compat_gates.py -q
225 passed, 1 skipped in 0.48s

Full backend unit suite on this branch:

6698 passed, 19 skipped

Two unrelated files (test_1081_physical_meter.py, test_subscription_auto_switch_pingpong.py) reported failures/errors in that local whole-suite run but pass on a pristine origin/dev worktree and pass in CI on both base and head — local suite-ordering artifacts on this host, not touched by this change (the diff is two test files). The four test_1771* hypothesis property files were excluded: the dependency is absent from this host.

Scope

Two test files, 18 insertions / 7 deletions. No production code.

[vybe]

Approving my own finding, with the reasoning stated so it can be checked rather than taken on trust.

This is a test-only reconciliation, not a behaviour change. The diff is two test files, +18/-7, no production code. Each of the four assertions was correct when written and became wrong when #1899 landed:

1. test_a_raising_check_is_logged read detail["skip_reason"], which stopped existing when the swallow moved from _skip to _fail. Asserting the FAIL shape is not a workaround — it is the property #1899 added, since only a counted status reaches hard_count/soft_count. The same file already had a passing sibling asserting "fail", so it was internally inconsistent.
2. test_persisted_check_error_does_not_replay_as_clean asserted soft_count == 0 for a persisted skipped+check_error row — it pinned the bug as expected behaviour. _did_not_pass counts that row deliberately. Flipped to 1 and kept with teeth: a row persisted by an older build still must not replay as clean.
   3/4. _build_local_template gained a required keyword-only is_bundled in #1899; both ent#89 call sites now pass it.

The process gap is the real finding. dev's four required checks run no unit tests, so auto-merge landed #1946 before regression diff reported. Two independently-green PRs can therefore redden dev on contact, with nothing blocking either. Worth considering regression diff as a required check — filing that separately rather than in this PR.

Verified: the three affected files → 225 passed / 1 skipped; full tests/unit on this branch → 6698 passed. The two unrelated files that misbehaved in my local whole-suite run pass on a pristine origin/dev worktree and pass in CI on both base and head, so they are local ordering artifacts, not regressions.