observability: phase 6.1 PR-C — 4 conformance fixtures (#24)

* observability: phase 6.1 PR-C conformance fillin

Drive four spec observability conformance fixtures end-to-end
through the existing test_observability harness:

- 002-otel-subgraph-hierarchy: §4.5 synthetic dispatch span
  parents under invocation; inner-node spans parent under
  dispatch span with the nested ["outer_sub", "inner_x"] /
  ["outer_sub", "inner_y"] namespace.
- 003-otel-error-status: §4.2 ERROR status mapping for the
  node_exception case (status_description == "node_exception",
  recorded exception event, openarmature.error.category
  attribute, ok_node stays OK).
- 007-otel-retry-attempt-spans: both sub-cases. Three sibling
  attempt spans with distinct attempt_index 0..2, single shared
  invocation parent. ERROR/ERROR/OK on retry-succeeds-third;
  ERROR x3 on retry-exhausts.
- 011-otel-determinism: signature-tuple comparison
  (name, status_code, sorted attrs minus the canonical
  non-deterministic-by-design ignore set: invocation_id +
  auto-generated correlation_id) across two runs.

Two minor YAML-shape translations kept inline in the
fixture-specific drivers so the existing adapter stays
unchanged (the adapter is shared with other suites; this PR
doesn't touch it):

- Fixture 007 flaky.fail_count=N rewritten to the adapter's
  failure_sequence shape before build_graph.
- Fixture 011 when:{field, gt: 0} edge syntax rewritten to the
  adapter's condition:{if_field, equals, then, else} shape via
  _translate_011_when_edges (gt:0 with the fixture's
  deterministic input is functionally equivalent to equals:1
  because counter==1 always under that input).

Map updates:

- _SUPPORTED_FIXTURES grows from 4 to 8.
- _DEFERRED_FIXTURES shrinks from 7 to 3 (004 / 006 / 010);
  each remaining note now spells out the specific gating-PR +
  spec/architectural rationale (PR-C.1 + proposal 0012 /
  PR-C.2 / PR-C.3) rather than the generic "deferred to
  Phase 6.1" placeholder.

Module docstring rewritten to distinguish the 8 driven
fixtures from the 3 deferred ones.

No engine, observer, or correlation changes. Pure test-side
work.

392 tests pass (was 388; net +4 from the new fixture drivers).
5 skipped (3 deferred conformance fixtures + 2 pre-existing).
Pyright clean.

* otel: address PR #24 review

- Fix invocation span status propagation. The OTelObserver was
  unconditionally calling set_status(OK) on the invocation span
  in _close_invocation_span; OTel doesn't auto-propagate child
  status to parents, so the spec §4.2 / fixture 003 contract
  ("invocation span ends ERROR when a child errored") wasn't
  satisfied. _handle_completed now sets ERROR on the invocation
  span when a child errors; _close_invocation_span no longer
  forces OK (clean invocations end UNSET, which exporters map
  to OK; ERROR is preserved by OTel SDK status precedence).
  Driver _run_fixture_003 gains the explicit invocation-status
  ERROR assertion + docstring fix.
- Fix _signature in _run_fixture_011 to include parent_name.
  Previously the docstring claimed "hierarchy" and the comment
  said "parent-by-name" but the signature ignored span.parent
  entirely. Now looks up parent.span_id in a per-run by-id map
  and includes the parent's name; a hierarchy regression where
  a node reparented to a different ancestor now surfaces as a
  signature divergence.
- Fix comment drift in _run_fixture_011_case: the translated
  edge condition is "equals: 1" (matching the deterministic
  counter == 1 input), not "equals: True".

Author: chris-colinsky

src/openarmature/observability/otel/observer.py

@@ -297,6 +297,17 @@ def _handle_completed(self, event: NodeEvent) -> None:
             span.set_status(Status(StatusCode.ERROR, description=event.error.category))
             span.record_exception(event.error)
             span.set_attribute("openarmature.error.category", event.error.category)
+            # Per spec §4.2 / fixture 003: the invocation span MUST
+            # end with ERROR status when any child node errors. OTel
+            # doesn't auto-propagate child status to parents — we set
+            # it explicitly here. The OTel SDK's status-precedence
+            # rule preserves ERROR through any subsequent
+            # ``set_status(OK)`` calls (only UNSET → OK transitions
+            # are honoured), so the close path's UNSET-leave still
+            # works for clean invocations.
+            inv_open = self._invocation_span.get(invocation_id)
+            if inv_open is not None:
+                inv_open.span.set_status(Status(StatusCode.ERROR, description=event.error.category))
         else:
             span.set_status(Status(StatusCode.OK))
         span.end()
@@ -839,11 +850,18 @@ def _close_invocation_span(self, invocation_id: str) -> None:
         open_span = self._invocation_span.pop(invocation_id, None)
         if open_span is None:
             return
-        # Status defaults to OK for completed invocations; if the
-        # engine surfaced an error, the failing node's span
-        # already carries it and OTel propagates ERROR up the
-        # parent chain on its own.
-        open_span.span.set_status(Status(StatusCode.OK))
+        # Don't unconditionally call ``set_status(OK)`` here. OTel
+        # doesn't auto-propagate child span status to parents, so
+        # the spec §4.2 / fixture 003 contract ("invocation span
+        # ends ERROR when a child errored") is satisfied by
+        # ``_handle_completed`` setting ERROR on this span when an
+        # error event fires. Calling ``set_status(OK)`` here would
+        # be a no-op when ERROR was already set (OTel SDK
+        # status-precedence preserves ERROR), but it's clearer to
+        # leave the status UNSET in the clean-completion path —
+        # exporters map UNSET to OK by convention, and the explicit
+        # ERROR-set in ``_handle_completed`` handles the failure
+        # path.
         open_span.span.end()
 
     def shutdown(self) -> None: