Activate fixture 037 case 5 (resume re-fire)

CHANGELOG.md

@@ -25,11 +25,16 @@ The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). The
 - **`observation.metadata.detached: true` moves to the parent-side dispatching observation** (proposal 0042, observability §8.4.2). The Langfuse mapping previously emitted `detached: true` on the dispatch observation inside the detached child trace; the §8.4.2 row added by 0042 places it on the **parent-side** dispatching observation that fires the detached child (the link observation in the main trace for detached subgraphs; the parent fan-out node observation for detached fan-outs). The detached-side observation no longer carries the flag.
 - **`LangfuseClient.update_trace` Protocol grows `input` / `output` keyword parameters** so observer-supplied values land on the Trace's headline fields.
 
+### Fixed
+
+- **`InvocationCompletedEvent.final_state` on the failure path now surfaces the partial state at failure point.** Spec §8.4.1 *Resume semantics* requires the failure-path `trace.output` hook to receive "the partial final state captured at the failure point"; the original PR #99 implementation defaulted to `starting_state`, so the hook saw pre-execution state when it should have seen post-execution-up-to-failure state. The engine now tracks the latest post-merge state via a `latest_state_box` on `_InvocationContext`, updated after every successful step and read on the failure path. Success-path behavior unchanged.
+- **`latest_state_box` is per-context, not shared across subgraph descents.** Unlike the sibling `final_node_box` (which shares by reference because the spec wants the innermost failing node's name — the real culprit), `latest_state_box` must isolate per level so the outermost Langfuse trace receives outer-state-typed values. Without the isolation, a subgraph-internal step's inner-typed state would leak up to the outer trace.output hook, breaking the hook's typed contract. Each subgraph / fan-out instance / parallel-branches branch gets its own fresh box. Pinned by four regression tests covering flat, subgraph, fan-out, and parallel-branches failure paths.
+
 ### Notes
 
 - **Pinned spec version bumped from v0.31.0 to v0.35.0.** Absorbs proposals 0042 (reserved-key extension), 0043 (Langfuse trace.input/output sourcing), and the textual additions in v0.32.0 (Gemini wire-format mapping, 0038, not yet implemented) and v0.33.0 (sessions capability, 0020, not yet implemented).
 - `LangfuseSDKAdapter` now applies `trace.input` / `trace.output` to the live Langfuse Trace. Input lands on the first real observation under the trace via `set_trace_io`; output uses a synthetic short-lived `openarmature.trace_io` observation as the carrier. The InMemoryLangfuseClient used by tests applies the fields directly.
-- Conformance fixture `observability/conformance/037-langfuse-trace-input-output` activated for the four decision-tree cases (default stub / `disable_state_payload=False` / hooks non-null / hooks null-fallthrough). Case 5 (resume re-fire) is deferred to a follow-up — needs the langfuse harness to grow checkpointer wiring + flaky-node test seam + two-phase multi-trace assertion.
+- Conformance fixture `observability/conformance/037-langfuse-trace-input-output` activated for all five cases (default stub / `disable_state_payload=False` / hooks non-null / hooks null-fallthrough / resume re-fire). The langfuse harness grew per-case `checkpointer: in_memory` wiring, a compact `flaky:` test seam, and a two-phase resume-flow assertion path.
 - The Langfuse v4 SDK marks `set_current_trace_io` / `Span.set_trace_io` deprecated ("removal in a future major version"). Empirical verification against Langfuse Cloud (v4.7.1, 2026-05-29) confirms it remains the **only** path that populates the Traces list view's headline `Input` / `Output` columns; `propagate_attributes(metadata=...)` does not substitute for it in the current UI. We will revisit when Langfuse publishes a concrete migration guide for v5.
 
 ## [0.10.0] — 2026-05-27

src/openarmature/graph/compiled.py

@@ -1059,20 +1059,38 @@ async def invoke(
             # box the engine populates as nodes enter; on the failure
             # path that's the inner-most node that raised, on the
             # success path that's the last node before the END-routing
-            # edge. ``final_state`` is the engine's returned state on
-            # success and ``starting_state`` on the failure path (the
-            # engine doesn't expose intermediate state across raises).
+            # edge. ``final_state`` precedence: the engine's returned
+            # state on success → the most recent successful step's
+            # post-merge state on a mid-graph raise (per §8.4.1
+            # *Resume semantics* "partial final state captured at the
+            # failure point") → ``starting_state`` only when no step
+            # ever completed.
             if context.final_node_box:
                 final_node = context.final_node_box[0]
             else:
                 # Defensive: invocation raised before any node fired
                 # (e.g., resume-path validation). Fall back to the
                 # declared entry node.
                 final_node = self.entry
+            # ``latest_state_box`` is typed ``list[Any]`` on
+            # _InvocationContext (the context isn't parameterized on
+            # StateT), but at the outermost level (where this code
+            # runs) it always holds an outer ``StateT`` from a
+            # successful step's post-merge state.  Cast for type
+            # narrowing; the per-context box-isolation pinned by
+            # ``test_failure_path_final_state_is_outer_type_*`` keeps
+            # this invariant honest.
+            event_final_state: StateT
+            if final_state is not None:
+                event_final_state = final_state
+            elif context.latest_state_box:
+                event_final_state = cast("StateT", context.latest_state_box[0])
+            else:
+                event_final_state = starting_state
             _dispatch(
                 context,
                 InvocationCompletedEvent(
-                    final_state=final_state if final_state is not None else starting_state,
+                    final_state=event_final_state,
                     status=status,
                     final_node=final_node,
                     invocation_id=invocation_id,
@@ -1201,6 +1219,15 @@ async def _invoke(
             else:
                 step_result = await self._step_function_node(node, current, state, context)
             state = step_result.state
+            # Proposal 0043 (post-PR-99 review): surface the most
+            # recent successful step's post-merge state so the
+            # outermost ``invoke()`` can populate
+            # ``InvocationCompletedEvent.final_state`` on the failure
+            # path with the partial state, not the bare initial state.
+            # Updated AFTER ``state = step_result.state`` so an
+            # exception inside the step bypasses this assignment and
+            # the previous value (or the empty box) survives.
+            context.latest_state_box[:] = [state]
 
             # Proposal 0043 (post-PR-99 review): restore the outer
             # ``current`` to the shared box after a successful step.

src/openarmature/graph/observer.py

@@ -497,6 +497,23 @@ class _InvocationContext:
     # descents so the inner-most node's name wins on failure (the
     # real culprit, not the wrapper).
     final_node_box: list[str] = field(default_factory=list[str])
+    # Per proposal 0043 (observability §8.4.1 *Resume semantics* +
+    # "partial final state captured at the failure point" clause).
+    # Tracks the most recent successful step's post-merge state at THIS
+    # context level so the outermost ``invoke()`` can populate
+    # ``InvocationCompletedEvent.final_state`` on the failure path with
+    # the partial outer state, not the bare ``starting_state``. On the
+    # success path the box is unused — the engine's return value is the
+    # canonical ``final_state``. **Distinct from ``final_node_box``**:
+    # the latest-state box is per-level (each subgraph / fan-out
+    # instance / parallel-branches branch gets its own fresh box),
+    # because the OUTER Langfuse trace cares about the outer-graph's
+    # state type, and an inner state has a different type. The
+    # ``final_node_box`` shares by reference because the spec wants the
+    # innermost failing node's name (the real culprit); state has the
+    # opposite contract — the outermost level's state is what the
+    # outer trace.output hook receives.
+    latest_state_box: list[Any] = field(default_factory=list[Any])
 
     def full_observers(self) -> tuple[SubscribedObserver, ...]:
         """Return the ordered observer list to deliver for events from
@@ -545,6 +562,9 @@ def descend_into_subgraph(
             drain_counters=self.drain_counters,
             state_cls=self.state_cls,
             final_node_box=self.final_node_box,
+            # latest_state_box is INTENTIONALLY NOT propagated — each
+            # context level tracks its own outer-state-typed latest
+            # successful step. See the field docstring above.
         )
 
     def descend_into_fan_out_instance(
@@ -596,6 +616,9 @@ def descend_into_fan_out_instance(
             drain_counters=self.drain_counters,
             state_cls=self.state_cls,
             final_node_box=self.final_node_box,
+            # latest_state_box is INTENTIONALLY NOT propagated — each
+            # context level tracks its own outer-state-typed latest
+            # successful step. See the field docstring above.
         )
 
     def descend_into_parallel_branch(
@@ -650,6 +673,9 @@ def descend_into_parallel_branch(
             drain_counters=self.drain_counters,
             state_cls=self.state_cls,
             final_node_box=self.final_node_box,
+            # latest_state_box is INTENTIONALLY NOT propagated — each
+            # context level tracks its own outer-state-typed latest
+            # successful step. See the field docstring above.
         )
 
     def take_step(self) -> int: