fix: preserve release Gate state through Part C

evals/gate/02-lifecycle.md

@@ -8,6 +8,11 @@ home**. Results never transfer between providers, between homes, or between
 candidates. A provider whose sequence is interrupted restarts at `L1` with a
 new clean home.
 
+`L1–L7` is the entry condition for that provider's Part C run. Keep the
+installed payload, `GATE_HOME`, and `GATE_CWD` in place while `G1–G6` run.
+`L8` is deferred until Part C completes or stops, so teardown cannot destroy
+the state that the live behavior cases must exercise.
+
 Part A must be complete, and `H1` must be authorized before any step that
 starts a provider session (`L7`, and any optional in-session check).
 
@@ -282,25 +287,28 @@ Requires `H1`.
   version and model id only. It does not prove the model will select the Skill
   for an arbitrary phrasing; routing behavior is out of scope for this Gate.
 
-## L8 — Teardown
+## L8 — Deferred teardown after Part C
 
-- **Setup.** `L7` recorded, or the sequence stopped.
+- **Setup.** If any case stops before `L7 PASS`, run teardown immediately.
+  Only an `L7 PASS` defers teardown until the corresponding provider's
+  `G1–G6` completes or stops.
 - **Action.** `run_skills_remove`; remove `GATE_HOME`, `GATE_CWD`, and
   any temporary manifest directory created by the session.
 - **Oracle.** No `run-with-first-tree` payload remains under either directory;
   the fixture directory digests are unchanged from `A3`.
 - **Evidence.** Final listings, the removal record, and the after-run fixture
   digests.
-- **Stop / fail.** A fixture digest changed during Part B — the Skill or the
-  session modified the source tickets. That is a `FAIL` for the run, not a
-  cleanup note.
+- **Stop / fail.** A fixture digest changed during Part B or Part C — the Skill
+  or a session modified the source tickets. That is a `FAIL` for the run, not
+  a cleanup note.
 - **Reset.** Not applicable; this is the reset.
 - **Claim limits.** Teardown covers local files only. Any First Tree chat
-  created during Part B is recorded, not deleted.
+  created during Part C is recorded, not deleted.
 
 ## Provider matrix
 
-Run every case above for both providers. The differences that matter:
+Run `L1`–`L7`, then the corresponding provider's Part C `G1`–`G6`, then `L8`
+for both providers. The differences that matter:
 
 | | Codex | Claude Code |
 | --- | --- | --- |

evals/gate/03-behavior-cases.md

@@ -1,8 +1,9 @@
 # Part C — Live behavior cases
 
 Part C is the only part of this Gate that qualifies live agent behavior. It
-requires `H1`, a completed Part A, and a passing Part B for the provider being
-exercised.
+requires `H1`, a completed Part A, and passing Part B cases `L1–L7` for the
+provider being exercised. Part B's `L8` remains deferred until this provider's
+`G1–G6` complete or stop.
 
 Run `G1`–`G6` **once per provider**. A case passed on Codex grants nothing to
 Claude Code and the reverse. Within a provider, run the cases in order; each
@@ -22,6 +23,11 @@ GATE_TMP=$(mktemp -d)      # operator-owned, separate from the session's own
 `GATE_HOME` and `GATE_CWD` keep the values Part B recorded for the provider
 under test.
 
+Start each provider session from `GATE_CWD` with `HOME="$GATE_HOME"`. Run all
+of `G1`–`G6` using the installed `PAYLOAD` that Part B recorded. Cases may
+start new sessions as their setup and reset instructions require; every such
+session reuses these Part B values.
+
 Copy [fixtures/](fixtures) from the clean candidate checkout to
 `GATE_FIXTURE_ROOT` once, record the per-file and directory digests from `A3`,
 and never edit them during the run.
@@ -485,8 +491,12 @@ session did with them.
 
 ## Provider notes
 
-- Run the whole of `G1`–`G6` in a session of the provider under test, using the
-  payload installed by that provider's Part B run.
+- Run the whole of `G1`–`G6` in sessions of the provider under test. Start each
+  from `GATE_CWD` with `HOME="$GATE_HOME"` and use the payload installed by that
+  provider's Part B run.
+- Run `L8` immediately after that provider's `G1`–`G6` completes or stops.
+  Teardown is still required when Part C fails, stops, or is inconclusive; it
+  does not change the Part C verdict.
 - The assignee agents in `G5` and `G6` may run on any provider; the case
   qualifies the **parent** session's provider. Record each assignee's runtime
   anyway, because an assignee's failure to answer makes the case

evals/gate/evidence-template.md

@@ -119,7 +119,7 @@ Fixture root (absolute, stable for the run):
 | `L5` rollback / recovery | | | | |
 | `L6` uninstall | | | | |
 | `L7` reinstall + fresh session | | | | |
-| `L8` teardown | | | | |
+| `L8` deferred teardown after Part C | | | | |
 
 ### Claude Code
 
@@ -132,7 +132,7 @@ Fixture root (absolute, stable for the run):
 | `L5` rollback / recovery | | | | |
 | `L6` uninstall | | | | |
 | `L7` reinstall + fresh session | | | | |
-| `L8` teardown | | | | |
+| `L8` deferred teardown after Part C | | | | |
 
 Exact removal command form that worked:
 

tests/test_release_gate.py

@@ -6,7 +6,9 @@
 
 ROOT = Path(__file__).resolve().parents[1]
 LIFECYCLE_GATE = ROOT / "evals" / "gate" / "02-lifecycle.md"
+BEHAVIOR_GATE = ROOT / "evals" / "gate" / "03-behavior-cases.md"
 RELEASE_GATE = ROOT / "evals" / "gate" / "04-release-smoke.md"
+EVIDENCE_TEMPLATE = ROOT / "evals" / "gate" / "evidence-template.md"
 SUPPORTED_AGENTS = ["codex", "claude-code"]
 
 
@@ -45,7 +47,9 @@ class ReleaseGateContractTests(unittest.TestCase):
     @classmethod
     def setUpClass(cls) -> None:
         cls.lifecycle = LIFECYCLE_GATE.read_text(encoding="utf-8")
+        cls.behavior = BEHAVIOR_GATE.read_text(encoding="utf-8")
         cls.release = RELEASE_GATE.read_text(encoding="utf-8")
+        cls.evidence = EVIDENCE_TEMPLATE.read_text(encoding="utf-8")
         cls.sections = {
             "Part B": cls.lifecycle,
             "D7": section(cls.release, "## D7 —", "## D8 —"),
@@ -103,6 +107,42 @@ def test_local_checkout_update_observation_is_not_overclaimed(self) -> None:
         self.assertIn("no installed skills found matching", update)
         self.assertIn("does not qualify a release update", update)
 
+    def test_part_c_uses_l1_through_l7_payload_before_deferred_l8(self) -> None:
+        lifecycle_intro = normalized_prose(
+            section(self.lifecycle, "# Part B —", "## Shared setup")
+        )
+        lifecycle_teardown = normalized_prose(
+            section(self.lifecycle, "## L8 —", "## Provider matrix")
+        )
+        behavior_intro = normalized_prose(
+            section(self.behavior, "# Part C —", "## Shared harness")
+        )
+        behavior_harness = section(
+            self.behavior, "## Shared harness", "## The confirmation contract"
+        )
+        provider_notes = normalized_prose(
+            self.behavior.split("## Provider notes", 1)[1]
+        )
+        d0 = section(self.release, "## D0 —", "## D1 —")
+
+        self.assertIn("l1–l7 is the entry condition", lifecycle_intro)
+        self.assertIn("keep the installed payload", lifecycle_intro)
+        self.assertIn("before l7 pass", lifecycle_teardown)
+        self.assertIn("only an l7 pass defers teardown", lifecycle_teardown)
+        self.assertIn("g1–g6 completes or stops", lifecycle_teardown)
+        self.assertIn("l1–l7", behavior_intro)
+        self.assertIn("l8 remains deferred", behavior_intro)
+        self.assertIn(
+            'Start each provider session from `GATE_CWD` with `HOME="$GATE_HOME"`.',
+            behavior_harness,
+        )
+        self.assertIn("run l8", provider_notes)
+        self.assertEqual(self.evidence.count("`L8` deferred teardown after Part C"), 2)
+        self.assertIn("| Part B `L1`–`L8` on Codex | `PASS` |", d0)
+        self.assertIn("| Part B `L1`–`L8` on Claude Code | `PASS` |", d0)
+        self.assertIn("| Part C `G1`–`G6` on Codex | `PASS` |", d0)
+        self.assertIn("| Part C `G1`–`G6` on Claude Code | `PASS` |", d0)
+
 
 if __name__ == "__main__":
     unittest.main()