Correct four release claims and fix two defects

[chris-colinsky]

Acts on spec's release review of v0.16.0..HEAD (coord release-v0.17.0/62), which found three things blocking the tag. This closes the first and takes the two code fixes spec recommended doing "while it is cheap". Blockers 2 and 3 are deliberately deferred, with reasoning at the bottom.

None of the four claims was a regression. Each was a statement about the release that was not true.

Four claims

The CHANGELOG was dated 2026-07-20. 67 days stale, and unchanged across ~70 commits including the one titled "Bring the CHANGELOG current before tagging". Now 2026-09-25, and RELEASING.md already requires re-confirming it on tag day.

The CHANGELOG contradicted itself about 0124. Line 40 said not-yet, line 61 said partial, the manifest said partial. The summary bullet predated the adoption that landed in HEAD.

0085 was implemented with half of it absent. Nothing writes enclosing_fan_out_lineage; compiled.py:372 says so in its own comment. Verified: zero write sites in src/. The consequence is narrower than "broken" and wider than "an optimization is missing":

• §10.11's no-mis-skip floor holds, because an unmatched record re-runs rather than skipping.
• §10.11.1's exactly-once guarantee does not extend to nested fan-outs, because a completed inner instance re-runs on resume and its side effects execute twice.

Fixture 076 passes against a record the test seeds by hand, not one the engine wrote. Now partial. The old note defended the gap by citing "§66 / §76", which do not exist in 0085; its actual Out of scope section lists three items and the write side is not among them.

0124's note named two unmet arms; there are three, and a fourth is retracted. Now listed worst first:

1. The resolver still keys on span openness, which §5.5 forbids outright, and no fixture can reach that check because the conformance adapter attaches instance or branch middleware where §5.1 specifies node middleware.
2. No subgraph span is synthesized on either observer, so an orphan inside a plain subgraph parents under the invocation span rather than the subgraph span.
3. The detached arms stay unmirrored on both observers.

The retracted one was a callable parallel branch on Langfuse. §8.4.8 makes that branch's single observation its per-branch dispatch observation, so there is no leaf-versus-dispatch race to lose, and the parent is stable with and without a yield (reproduced both ways). It reached a shipped manifest note from a review finding that was never reproduced. The residual callable-branch gap is real but narrower: that observation omits parallel_branches_parent_node_name, which §8.4.2 requires. Tracked, not fixed here.

Two defects, both public API

StructuredOutputInvalid used the wrong field names. It carried raw_content and failure_description where llm-provider §7 names output_content and error_message, and §7.1 says the reask builder receives those two. So a caller writing the documented builder hit AttributeError, openai.py:803 converted it to raise exc from reask_error, and the call failed on the first invalid output exactly as if no builder had been supplied. Reask did not work as documented.

Renamed across 62 references in 7 files. This is a pre-1.0 break for a caller reading either attribute. The conformance harness carried an alias bridging the two spellings, with a comment saying it would stay correct if the implementation ever renamed to match; it has, so the alias resolved to itself and is removed.

A per-attempt retry override replaced the extras container. §7.1: "undeclared extras (§6) merge by the same per-key rule." The analogue of a declared field is an extras KEY rather than the container, so a key the override sets replaces that key and a key it does not mention inherits the base's. The implementation replaced the whole container and logged the base keys it dropped, so an override adjusting one vendor knob silently withdrew every other knob the caller had set for that attempt. Nothing is dropped now, so the warning goes with it.

test_per_attempt_override_with_extras_replaces_and_warns asserted the defect. It is inverted and renamed, and now asserts the inherited keys, which is the half that can fail: the override's own key arrives under either reading, so asserting only that would pass against the old behaviour.

Evidence

Mutation	Result
extras reverted to replace-the-container	killed (the inverted test)
harness carries-lookup returns None	killed (conformance 022, 023)

2301 passing, ruff and pyright clean, manifest consistent.

Deferred, and why

Blocker 2 (the resolver's openness check). Pre-existing: orphan parentage depended on scheduling before 0124 too. 0124 already ships partial, and the note now names this as the primary unmet arm rather than a footnote. The fix also needs the adapter attachment changed before any fixture can hold it, which makes it two pieces of work rather than one.

Blocker 3 (fixtures that run and assert nothing, and extending the 0120 vocabulary check one level down into expected: / resume: / nodes:). Spec calls this the highest-leverage item in the review and I agree. It is also not a regression, it is substantial, and the last three review rounds on #302 showed what happens when work of this size is done under tag pressure. It wants its own pass.

Both are tracked locally with spec's measurements attached.

Also noted, not fixed

scripts/check_conformance_manifest.py validates that every accepted proposal has an entry and that status is one of four strings. Nothing compares a status to the code, which is how 0085 shipped implemented over a stub and passed cleanly. That is the same green-is-the-null-result failure AGENTS.md warns about for fixtures, one level up. Worth a guard; not one to build under tag pressure.

[Copilot]

Copilot review overview

🟡 Changes recommended

Unresolved changelog, manifest, documentation, and test-helper inconsistencies remain.

Review effort: Lite
Findings: 3

Open (3)

• Correct article usage before error_message · New
• Update retry extras docstrings for per-key merging · New
• Update assert_error_carries docs to renamed fields · New

What changed in this PR

Corrects v0.17.0 release claims and fixes structured-output error naming and retry extras merging.

Changes:

• Aligns release metadata and proposal statuses.
• Renames structured-output error fields to the documented API.
• Merges retry extras per key and updates tests and documentation.

File	Summary
tests/​unit/​test_structured_output.py	Updates validation assertions for renamed fields.
tests/​unit/​test_observability_otel.py	Updates reask assertions; simplify the obsolete identity alias map. Nit (1 vote).
tests/​unit/​test_llm_provider.py	Strengthens retry extras regression coverage.
tests/​conformance/​harness/​wire.py	Removes legacy aliasing; update the carry assertion docstring. Nit (4 votes).
src/​openarmature/​llm/​retry.py	Updates retry terminology; document per-key extras merging. Nit (1 vote).
src/​openarmature/​llm/​providers/​openai.py	Applies renamed fields and per-key extras merging; update stale docstrings. Nit (3 votes).
src/​openarmature/​llm/​errors.py	Renames public exception fields.
docs/​concepts/​llms.md	Updates API examples; correct grammar and describe per-key merging. Nits (4 and 1 votes).
conformance.toml	Corrects proposal statuses; update the stale alias note. Nit (1 vote).
CHANGELOG.md	Updates release claims; resolve contradictory status, reask, and extras-merging notes. Nits (1 vote each).

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.