docs(adr-013): Erratum E3 — lift the pipeline_core_version caveat, in part

[Polichinel]

Closes #228, with one correction to what it asked for.

What #228 asked

§2.2 declared pipeline_core_version unreliable until pipeline-core cut a real release. 3.0.0 shipped to PyPI on 2026-08-03, so the condition is met. Lift the caveat.

Agreed — but the suggested wording states one half as current when it is not.

The correction: released versus merged

producer installed as	pipeline_core_version	in force?
a released wheel	authoritative — the version was written by the release that built it	yes, now
an editable install	"unknown"	no — not in any released version

pipeline-core#403 makes the editable case report "unknown" rather than a stale number. It merged 2026-08-04 13:58. 3.0.0 was uploaded 2026-08-03 02:06 — a day and a half earlier. So the fix is on pipeline-core's main and in no release. It becomes true of producers at their next one.

#228's suggested note reads as though both halves are live. E3 says which is which.

Verified here, not taken on trust

importlib.metadata       -> 2.3.0
source beside it         -> 3.0.0
_pipeline_core_version() -> 'unknown'

A full major version stale — exactly the value that would have been stamped into published provenance, and exactly what the fix suppresses.

The operational point, which #228 got right

Consumers must treat "unknown" as "do not infer the producing version", never as an error — and the set of runs producing it will widen at pipeline-core's next release, because every developer run joins it. A consumer that starts rejecting "unknown" on the strength of this lift breaks exactly those runs.

Nothing in this repository changes

Hop-B re-embeds the Hop-A header untouched, so we forward this value and never compose it. contract_version stays 1.5.

The claim carries a check, because it expires

E3 says the fix is unreleased. That stops being true the day pipeline-core publishes again, and nothing here would signal it.

test_e3s_claim_about_unreleased_behaviour_is_still_true fails if the installed pipeline-core is a released distribution past 3.0.0. Inert in a maintainer's editable environment, live in CI which installs from PyPI — stated in the docstring rather than left to be discovered, since a guard that only runs in one place is half a guard. Mutation-proven against a synthetic 3.1.0 wheel.

Downstream

#228 notes views-faoapi is the other reader of this field and inherits both halves. Worth a note to them; not done in this PR.

ruff check .    # clean
pytest -q       # 399 passed, 1 skipped, 40 xfailed

🤖 Generated with Claude Code