Summary

PR-C of Phase 6.1: drive four spec observability conformance
fixtures end-to-end through the existing
test_observability harness. Pure test-side work — no
engine, observer, or correlation changes.

Driven (this PR)

• 002-otel-subgraph-hierarchy — §4.5 synthetic subgraph
  dispatch span parents under the invocation; inner-node spans
  parent under the 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×3 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 of the
  same graph.

Deferred to follow-up PRs

Each remaining _DEFERRED_FIXTURES entry now spells out the
specific gating-PR + spec/architectural rationale rather than
the generic "deferred to Phase 6.1" placeholder:

• 004-otel-routing-error-attribution → PR-C.1 (gated on
  proposal 0012's spec amendment + v0.9.0 spec submodule
  bump). Routing-error attribution requires the §3/§6 ordering
  swap so RoutingError lands on the preceding node's
  completed event with error populated.
• 006-otel-fan-out-instance-attribution → PR-C.2.
  Spec §5.4 mandates per-instance dispatch spans nested between
  the fan-out node span and the inner-node spans, plus three
  fan-out-config attributes the observer doesn't currently
  surface (item_count, concurrency, error_policy).
  Current observer opens one shared synthetic span at the
  namespace prefix instead.
• 010-otel-log-correlation → PR-C.3.
  Needs synchronous observer prep hook (prepare_sync) so
  the engine task can attach the observer's span to OTel
  context for the duration of node-body execution. Observer
  span creation runs on the worker task today and isn't
  available synchronously after _dispatch_started.

PR-C.1 / PR-C.2 / PR-C.3 are mutually independent and land in
any order once their respective gates clear (PR-C.2 + PR-C.3
have no gate; PR-C.1 waits on the spec submodule bump).

Two minor YAML translations kept inline

Both fixture-specific so the shared adapter stays unchanged
(the adapter is used by other test 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: counter, gt: 0} edge syntax
  rewritten to the adapter's
  condition: {if_field, equals, then, else} shape via
  _translate_011_when_edges (the deterministic input means
  counter == 1 always, so gt: 0 is functionally
  equivalent to equals: 1 for this fixture's flow).

Spec agent flagged the harness-DSL gap as a small non-blocking
follow-up — extend tests/conformance/adapter.py edge-condition
DSL with gt / gte / lt / lte / in / not_in
so future fixtures with arithmetic edge conditions don't need
per-driver translation. Tracked separately under harness backlog
in docs/phase-6-1-conformance-fillin.md.

Test plan

• uv run pytest -q — 392 passed (was 388; net +4 from
  the new fixture drivers), 5 skipped (3 deferred
  conformance fixtures + 2 pre-existing)
• uv run pyright — 0 errors
• uv run ruff check . && uv run ruff format — clean
• No engine/observer/correlation changes — diff is one
  file (tests/conformance/test_observability.py)