test(ci): wire 17 uncollected infra/libs tests into CI and name the second edit in the concurrency guard (#15051, #15302)

[mrveiss]

Thinking Path

#15051 re-measured before touching anything, because the issue's own comment said its table predated PR #15048. Static count against current Dev_new_gui:

module	test functions	marked (integration/slow/distributed/performance)	unmarked
test_architecture_compliance.py	15	2	13
test_db_initialization.py	6	6	0
test_distributed_system_integration.py	6	6 (gained module-level pytestmark via #14979/#15166, after the issue was filed)	0
test_performance_optimization.py	5	5 (same #14979/#15166 change)	0
test_redis_db_ssot.py	4	0	4
libs/.../test_integration.py	5	5	0
total	41	24	17

Of the original 28 unmarked, 11 (the whole test_distributed_system_integration.py and test_performance_optimization.py populations) picked up a marker in an unrelated PR (#15166, merged after #15051 was filed and after the #15048 follow-up comment) and now run via marker-tests.yml's third invocation. 17 remain — 13 in test_architecture_compliance.py, 4 in test_redis_db_ssot.py — and those ran in no workflow and no bare pytest, ever.

The dead ci.yml step (Run integration tests, shard 1 only) named infrastructure/shared/tests/integration, a path #734 renamed to autobot-infrastructure/shared/tests/integration without updating this reference. It quietly succeeds: if [ -d "$TEST_DIR" ]; then ...; else echo "::warning::..."; fi has no else exit 1, so the step always exits 0 and has done since #734 — not an error, a silent no-op green on every shard-1 run.

Both defects are fixed, not just documented:

1. Wiring — autobot-infrastructure/shared/tests and libs added to pytest.ini testpaths and to all four repo_tests-carrying pytest invocations (ci.yml, coverage.yml, test-durations.yml, marker-tests.yml's first invocation), so repo_tests/hook_suites_run_in_ci_test.py's root-set agreement stays true and the 17 unmarked tests run in the PR gate for the first time.
2. Dead step — retired with a reason recorded inline, not repointed. Repointing at the now-correct path would have duplicated collection of .../tests/integration (already reached by the wired-in invocation above) and put its integration-marked tests into the required, blocking PR gate with no -m filter of its own — the exact "separate decision with its own blast radius" marker-tests.yml's own comment declines to make inline.

Wiring in surfaced two real bugs, fixed on the implementation side per the "fix, don't adjust assertions" rule:

• get_backend_config() never set a "host"/"port" key — only "server_host"/"server_port" (the bind address). utils/service_discovery.py:236's _register_backend_service already reads backend_config.get("host") and logs an error on every process start when it comes back None, silently falling back to system_defaults. Fixed by wiring in the existing canonical get_host("backend")/get_port("backend") resolvers (env var → infrastructure.hosts.backend → fallback map), the same precedence every sibling accessor here already uses.
• test_no_hardcoded_ips_in_redis_helper imported from utils.redis_helper import REDIS_HOST — utils/redis_helper.py was deleted from every tree years ago (confirmed via git log --diff-filter=D), and REDIS_HOST was never a real export of it even historically. Repointed at the canonical autobot_shared.redis_client accessor, mirroring the identical fix ci(gates): make four gates that could not fail actually fail (#13543, #13286, #13200) #15048 already made for the sibling TIMEOUT_CONFIG import in the neighboring test.

Also rewrote test_service_discovery_has_defaults: it asserted four keys inside a service_discovery_defaults config section that two call sites (service_discovery.py, distributed_service_discovery.py) already document inline as having "no ssot equivalent". Nothing writes those keys any more post-#13286's SSOT consolidation — the test is now provably wrong, not the implementation, so it now asserts the property those two callers actually depend on: the section resolves to {}, never None.

The VM-topology comparison tests (test_redis_on_vm3_only, test_frontend_on_vm1, etc.) were reasoned through from source as far as static reading allows without executing code (forbidden by task constraints) — several compare a config accessor's output against the same NetworkConstants source it was built from, so they hold by construction. test_redis_on_vm3_only compares against a 5-tier ConfigRegistry resolution chain that cannot be fully verified without running pytest; left as-is for CI to be the actual judge, per the task's own instruction, and will be triaged from real CI output if it turns up red.

#15302: the guard's assertion at :104-108 told an author their omission was wrong but never mentioned DELIBERATELY_EXEMPT — exactly how #15300 got missed. Chose the message route over coupling the two edits: the file's own docstring already weighed this trade-off ("the coupling route is stronger but needs a way to recognise a required-context shim, which is a judgement rather than a pattern") and a judgement-based coupling risks false positives against this repo's preference for mechanical, derived checks over heuristics. The assertion now names the exemption dict, the exact filename, and cites #15300 as the precedent.

What Changed

• pytest.ini — testpaths gains autobot-infrastructure/shared/tests and libs.
• .github/workflows/ci.yml — same two roots added to the PR-gate pytest invocation; the dead Run integration tests step is retired with its reasoning recorded inline.
• .github/workflows/coverage.yml, .github/workflows/test-durations.yml, .github/workflows/marker-tests.yml — same two roots added to keep the four repo_tests-carrying invocations in agreement.
• autobot-backend/config/service_config.py — get_backend_config() now sets "host"/"port" via the canonical get_host/get_port resolvers.
• autobot-infrastructure/shared/tests/integration/test_architecture_compliance.py — test_no_hardcoded_ips_in_redis_helper repointed off a deleted module; test_service_discovery_has_defaults rewritten against the current SSOT reality.
• repo_tests/workflow_concurrency_guard_test.py — the concurrency-guard failure message now names DELIBERATELY_EXEMPT as the second edit.
• repo_tests/infra_libs_test_wiring_guard_15051_test.py (new) — the AC Dev new gui #5 guard: a tracked test module under autobot-infrastructure/shared/tests or libs must be named by pytest.ini or some CI invocation, marked or not. Scoped to these two roots (not repo-wide) with a reach floor of 5 modules.

Verification

Execution is off-limits for this task (do not run pytest / execute code from the codebase) — reasoning is from source, cross-checked against:

• Ran this PR's own new guard's discovery logic by hand against the current tree (git ls-files / regex extraction only, no pytest) and confirmed zero orphans under autobot-infrastructure/shared/tests and libs, and 67 pre-existing orphans elsewhere in the repo that a repo-wide version of the same check would have wrongly reported against this PR.
• git log --diff-filter=D confirming utils/redis_helper.py's deletion and its historical contents (no REDIS_HOST export ever existed).
• service_discovery.py:236's existing .get("host") read and its logger.error fallback, proving the missing key is a real, already-manifesting defect.
• Inline comments in service_discovery.py / distributed_service_discovery.py documenting service_discovery_defaults as having no SSOT equivalent.
• python3 -c "import ast; ast.parse(...)" on every touched .py file and yaml.safe_load on every touched workflow — all parse cleanly.
• repo_tests/hook_suites_run_in_ci_test.py and repo_tests/marker_suite_root_coverage_test.py read in full to confirm the exact root-set-agreement and floor mechanics the new roots must satisfy, and that none of the existing floors (MARKER_MIN_PASSED=34, infra=10, slm=0) are lowered or bypassed — only more roots are added to invocations that already clear those floors.
• CI on this PR is the actual judge of the 17 newly-collected tests and the two provisional fixes; any red result will be triaged from real output rather than guessed at further.

Model Used

Claude Sonnet 5

Refs #15051 -- wiring, dead-step retirement, and the two implementation bugs it surfaced are complete; the VM-topology assertions in test_architecture_compliance.py could not be verified without executing pytest (out of scope for this task) and are left for CI to judge, which is why this is Refs rather than Closes.
Refs #15302 -- the message-route fix for the guard's third, previously-unmet criterion is complete; left as Refs rather than Closes per this repository's closing convention (closed by hand, with evidence, not by a merge keyword).

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No new hardcoded values of either class — ssot and other both block.

Known backlog in pipeline-scripts/hardcoded_values_baseline.txt is suppressed and tracked in #14371.

[mrveiss]

CI found a real defect — in the guard, not in the wiring. Fixed in 5acf4a80f5.

python-suite shard 7/12 failed on pipeline-scripts/pytest_root_collection_floor_test.py::TestRootDerivation::test_the_derivation_agrees_with_an_independent_yaml_parse:

At index 0 diff: 'autobot-backend' != 'libs'

Both lists held the same thirteen roots. Only libs' position differed — and it came from prose.

That check cross-checks two parses of marker-tests.yml. The script anchors on the literal three-token python -m pytest, which no comment produces. The test did tokens.index("pytest") over the whole run: block and started reading roots from there. This PR added a comment explaining the wiring, and that comment contains the word pytest and names the sibling step "Run marked tests — infrastructure, libs and frontend". So the test's parse began inside the comment and read libs — a real directory, so it passed the existence filter — as the first root, out of English.

The workflow was correct the whole time. The check went red against correct configuration, which is the failure mode that teaches people to ignore a check.

The fix

_command_tokens strips shell comments before tokenising, and _invocation_starts anchors on the full python -m pytest sequence rather than a bare pytest token — matching what the script it cross-checks has always done. Two regression tests pin it:

• test_a_comment_naming_pytest_does_not_become_a_root — the exact shape above; asserts libs does not survive the strip and the invocation anchors at the command.
• test_a_pip_install_line_is_not_read_as_an_invocation — a latent second bug the same change closes: step 2's pip install pytest pytest-asyncio … matched the old bare-token search too. It happened to name no existing paths, so it never produced a wrong root — it was one dependency rename away from doing so.

Verified

Recomputed both sides as data (no repo code executed): the new parse yields

autobot-backend, autobot_shared, autobot-tts-worker, repo_tests, tools, scripts,
pipeline-scripts, autobot-infrastructure/shared/scripts/hooks, .claude/skills/claims-audit,
autobot-infrastructure/shared/tests, libs, autobot-frontend/tests, autobot-slm-backend

which matches workflow_roots' output from the failing CI run exactly, in order.

Worth stating plainly: none of the 17 newly-collected tests failed. Every one of the ten required contexts was already green, and python-suite is not a required check — so this could have been merged past. It should not have been: the entire point of this PR is that a test which never runs proves nothing, and merging over a red result from the suite it just wired in would have reproduced that defect in a new place on the same day.

[mrveiss]

Triage of the remaining failures.

Run slow / integration / distributed / performance tests fails on autobot-backend/system_benchmarks_performance_test.py::test_system_startup_performance. This is pre-existing on base, not introduced here — marker-tests.yml run on Dev_new_gui (2026-08-28) fails the same test with the same assertion, and this PR does not touch that file or the code path it measures. Evidence and analysis posted to #15055, including that the assertion's claimed load-invariance does not hold: 700 work units on base versus 3451 here, with the calibration absorbing only 1.9x of a 4.9x swing, on a run that changed no relevant code.

The four test_architecture_compliance.py failures were real and are now marked, with reasons, as #15194's — see the commit and the writeup on that issue. They assert a fixed VM topology the platform does not have; the owner has confirmed AutoBot runs in Docker, on one VM, or on any number chosen. I did not edit them green, which would have hidden the defect this PR exists to expose.