docs(tests): clarify test_inner_create_fail_refunds_in_creation_tx docstring

Expand the docstring to explicitly reference the geth bal-devnet-3
behavior the test codifies, spell out the regression signal, and
justify the `outer_halts` omission by pointing to PR ethereum#2689's
`test_creation_tx_failure_preserves_intrinsic_state_gas`.

Author: spencer-tb

tests/amsterdam/eip8037_state_creation_gas_cost_increase/test_state_gas_create.py

@@ -2115,28 +2115,31 @@ def test_inner_create_fail_refunds_in_creation_tx(
 
     The original closed PR #2639 test asserted inner CREATE's
     `GAS_NEW_ACCOUNT` PERSISTS on child failure (targeting a
-    bal-devnet-3 geth behavior). PR #2704 subsequently changed the
-    spec to refund that charge, inverting the original premise.
-    This test covers the CURRENT spec behavior in a scenario not
-    exercised by PR #2704's own tests (which all use non-creation
-    factories):
+    bal-devnet-3 geth behavior where geth incorrectly refunded).
+    PR #2704 later changed the spec to refund the charge, which
+    codifies what geth was already doing and inverts the original
+    test's premise. This test covers the CURRENT spec behavior in
+    a scenario not exercised by PR #2704's own tests (which all
+    use non-creation factories):
 
     A creation tx (to=None) whose initcode performs `num_inner_ops`
     inner CREATE/CREATE2 calls with REVERT initcode. Each inner
     CREATE charges `GAS_NEW_ACCOUNT` state then refunds it via the
     child-error branch (PR #2704). The outer initcode then
     terminates via RETURN (succeeds) or REVERT.
 
-    Both outcomes yield `block_state = outer intrinsic` because
-    inner refunds net the state gas back to zero; PR #2689's top
-    level refund (applied on revert) is a no-op over an already
-    zeroed `state_gas_used`.
+    Both outcomes yield `block_state = outer intrinsic` because inner
+    refunds net the state gas back to zero; PR #2689's top-level
+    refund (applied on revert) is a no-op over an already-zeroed
+    `state_gas_used`. A client regressing to the pre-#2704 "persist"
+    behavior would inflate `block_state` by
+    `num_inner_ops * GAS_NEW_ACCOUNT` and fail both variants.
 
-    A client that regresses to the pre-#2704 "persist" behavior
-    would inflate `block_state` by `num_inner_ops * GAS_NEW_ACCOUNT`.
-    The `outer_halts` variant is omitted because `INVALID` absorbs
-    all remaining gas into `regular_gas_used`, letting `block_regular`
-    dominate the header and dilute the state dimension signal.
+    The `outer_halts` variant is omitted: INVALID absorbs remaining
+    gas into `regular_gas_used`, making `block_regular` dominate the
+    header and dilute the state-dimension signal. PR #2689's
+    `test_creation_tx_failure_preserves_intrinsic_state_gas` already
+    covers the creation-tx + top-level halt interaction.
     """
     gas_costs = fork.gas_costs()
     outer_state_gas = fork.create_state_gas(code_size=0)
@@ -2158,10 +2161,8 @@ def test_inner_create_fail_refunds_in_creation_tx(
 
     if outer_outcome == "succeeds":
         termination = Op.RETURN(0, 0)
-    elif outer_outcome == "reverts":
-        termination = Op.REVERT(0, 0)
     else:
-        termination = Op.INVALID
+        termination = Op.REVERT(0, 0)
 
     initcode = setup + inner_ops + termination