docs: finalize bug_baseline_phase_test_isolation (PR #466 merged)

docs/00_overview/planned_features/02_mvp2/bug_baseline_phase_test_isolation/implementation_plan.md → docs/00_overview/implemented_features/2026_06_05_bug_baseline_phase_test_isolation/implementation_plan.md

@@ -1,7 +1,7 @@
 # Implementation Plan — Baseline-phase unit tests depend on suite ordering
 
 **Date:** 2026-06-02
-**Status:** Ready for Execution
+**Status:** Complete (PR #466, squash-merged `6298e77`, 2026-06-05)
 **Primary spec:** [feature_spec.md](feature_spec.md)
 **Policy source(s):** [CLAUDE.md §"Bug Fix Protocol"](../../../../../CLAUDE.md), [docs/05_quality/testing.md](../../../../05_quality/testing.md)
 

docs/00_overview/implemented_features/2026_06_05_bug_baseline_phase_test_isolation/pipeline_status.md

@@ -0,0 +1,25 @@
+# Pipeline Status — Baseline-phase unit tests depend on suite ordering
+
+## Idea
+- Status: Complete
+- File: idea.md
+
+## Spec
+- Status: Approved
+- Date: 2026-06-02
+- File: feature_spec.md
+- Cross-model review: GPT-5.5 passed (2 cycles, 2 findings — 2 accepted, both Low)
+- Phases: 1 total, 1 covered by spec (single-phase)
+
+## Plan
+- Status: Approved
+- Date: 2026-06-02
+- File: implementation_plan.md
+- Cross-model review: GPT-5.5 passed (2 cycles, 1 finding — rejected-with-counter-evidence, Low)
+- Stories: 2 total across 1 epic
+- Phases covered: 1 of 1 (single-phase)
+
+## Implementation
+- Status: Complete (PR #466, squash-merged `6298e77`, 2026-06-05)
+- Release: mvp2
+- Note: Backend test-only, no migration. Story 1.1 deferred the `get_settings()` call in `_compute_baseline_wait_s` (`backend/workers/orchestrator.py`) into the falsy-`trial_timeout_s` branch so explicit-timeout callers never construct `Settings` (return values unchanged). Story 1.2 added an autouse `_settings_env_and_restore` fixture + a `test_explicit_timeout_does_not_read_settings` regression to `test_orchestrator_baseline_phase.py`. Standalone run with secrets unset: 14 passed (pre-fix `3 failed, 1 passed`); full unit suite 2400 passed. No Gemini findings; final GPT-5.5 skipped (≤40 LOC, test-only, below threshold). All 19 CI checks green.

state_history.md

@@ -4,6 +4,16 @@
 
 ---
 
+### `bug_baseline_phase_test_isolation` — hermetic baseline-wait unit tests via lazy settings read (PR #466, 2026-06-05)
+
+**What shipped.** `backend/workers/orchestrator.py`'s `_compute_baseline_wait_s(study)` computed `min(600, max(60, trial_timeout_s + 30))`, reading `study.config["trial_timeout_s"]` or falling back to `get_settings().studies_default_timeout_s`. The bug: it called `get_settings()` **unconditionally** at the top, even when an explicit `trial_timeout_s` was present and the settings value was never used. `Settings` construction requires `DATABASE_URL_FILE` + `POSTGRES_PASSWORD_FILE` (CLAUDE.md Rule #2), so the three explicit-timeout `TestComputeBaselineWaitS` cases only passed when an *earlier* test module in the same session had already seeded those env vars into the process — they failed standalone (`3 failed, 1 passed`) with the secrets unset. **Backend test-only — no migration, no API, no runtime behavior change** (Alembic head stays `0023`). 2 stories / 1 epic.
+
+**The fix (both layers, per Bug Fix Protocol "fix at the right layer").** Story 1.1 (FR-1) defers the `get_settings()` read into the missing/falsy-`trial_timeout_s` branch — `trial_timeout_s = study.config.get("trial_timeout_s"); if not trial_timeout_s: trial_timeout_s = get_settings().studies_default_timeout_s` — so explicit-timeout callers never construct `Settings`. The falsy-fallback semantics (`None`/`0`/missing all fall back) and the formula constants are byte-for-byte preserved; return values are unchanged for every input. Story 1.2 (FR-2/FR-3) adds an autouse `_settings_env_and_restore` fixture (mirrors `test_poll_cron_kwargs.py`) that seeds both secret env vars at `/dev/null` and clears the **canonical** `get_settings` lru_cache imported from `backend.app.core.settings` — NOT `orch.get_settings`, which `test_missing_trial_timeout_uses_settings_default` monkeypatches to a plain lambda with no `.cache_clear()` — making the module hermetic regardless of collection order. It also adds `test_explicit_timeout_does_not_read_settings`, which monkeypatches `orch.get_settings` to a `_boom()` that raises, asserting the explicit-timeout path never calls it.
+
+**Verification.** Reproduced the failure first: `env -u DATABASE_URL_FILE -u POSTGRES_PASSWORD_FILE pytest test_orchestrator_baseline_phase.py -p no:randomly` → `3 failed, 1 passed` pre-fix. Post-fix: **14 passed** standalone with secrets unset (AC-1/AC-3). The new regression test was confirmed to **fail** against the pre-fix production code (stashed Story 1.1 → `1 failed`) and passes post-fix (AC-2). Full unit suite `make test-unit`: 2400 passed, 1 skipped (AC-4, coverage gate satisfied). `make lint` 0 errors + `mypy --strict` clean on both touched files. No Gemini findings; final GPT-5.5 review skipped (≤40 LOC, test-only, no studies/judgments/adapter/migration surface — below the threshold). All 19 `pr.yml` checks green.
+
+**Context.** Iteration 3 of the operator's "repeat this process three more times" run. Selected after the user chose "next locally-verifiable item" over the mechanically-top backlog item: the priority-sorted top ungated plan-ready item was `chore_demo_seeding_integration_tests_rewrite` (a 19-story DB-only integration-test rewrite unverifiable in this no-Docker/no-Postgres environment → CI-only), and the two `feat_` siblings above it are product-gated (see below). `bug_baseline_phase_test_isolation` is the highest-priority item whose tests run locally without a database. **Drift correction made this iteration:** iteration-2's finalization had described the two normalizer siblings (`feat_apply_path_normalizer_declaration`, `feat_query_normalizer_typed_pipeline`) as "now UNGATED" after PR #459; that overstated — PR #459 only cleared their G-1 *dependency* gate. Both retain **product gates** (apply-path: G-2 operator-friction evidence; typed-pipeline: open Q-1 Product decision + a "Do NOT `/impl-execute`" design-ahead banner) and are NOT auto-executable. `state.md`'s plan-ready line was corrected accordingly.
+
 ### `chore_cluster_detail_rung_badge` — cluster-detail UBI readiness card with rung badge (PR #464, 2026-06-05)
 
 **What shipped.** A new `ClusterDetailUbiReadinessCard`, unconditionally mounted on `/clusters/[id]` between `<ClusterActionBar>` and `<ClusterDetailIndicesCard>`, so an operator can read a cluster's UBI readiness rung directly on the detail page instead of having to open the generate-judgments dialog (the only prior surface for `<UbiRungBadge>`). **Frontend-only — no backend, no new endpoint, no migration** (Alembic head stays `0023`). 8 stories / 1 epic, executed 8 → 1 → 2 → … → 7 (Story 8 first so the shared-hook change lands before the AC-8 assertions depend on it).