yaml: every extracted step is named run — func_start captures no name, so orphan/duplicate census and per-step identity are all dead #2767

Same pass as #2763–#2766 (2026-09-05, engine 3b3568e0, corpus main 6183c4b). raw_state_slop_orphans is the corpus's third-weakest metric (61% consistency) after #2765's two, and yaml sits at −100% on it. The cause is upstream of the orphan census: yaml's thirteen extracted functions all carry the same name.

Measured

galaxyscope --db-only over keyword-rosetta/data/yaml, function_data joined to file_data:

a.yml     run   run   run
b.yml     run   run   run
c.yml     run   run   run
main.yml  run   run   run   run

Thirteen slices, one name. yaml.py's func_start has no capture group —

"func_start": re.compile(r"^[ \t]*(?:-?[ \t]*run:|script:|before_script:|after_script:)[ \t]*[|>]*")

— so _extract_name falls back to the matched keyword. The literal-capture helper that builds _keyword_bucket_names (detector.py:1010) does not recover run from this pattern (the -?[ \t]* prefix defeats it), which is confirmable directly:

>>> _closed_literal_capture(LANGUAGE_DEFINITIONS['yaml']['rules']['func_start'].pattern)
frozenset()          # css → {'@media','@keyframes',...}; dockerfile → {'RUN','CMD',...}

So unlike css's keyframes and dockerfile's RUN — which #2728 deliberately excluded from the orphan and duplicate censuses precisely because a keyword cannot be orphaned — yaml's run reaches both censuses as if it were an identifier, and is then dropped by the len(func_name) > 3 guard on detector.py:1496. Two different mechanisms both silently discarding the same slice.

What is lost

• Orphan census: raw_state_slop_orphans 0.00 against a 2.50 median, and risk_tech_debt inherits it.
• Duplicate census: thirteen same-named slices with different bodies. The body_hash guard (duplicate_logic heuristic is scope-blind: same-named shadowed local helpers across different where-blocks flagged as duplicates #1498) is the only thing preventing thirteen false duplicates — the name test alone would have fired on every one.
• Per-step identity, which is the real cost outside the corpus. In a scanned repo, every step of every workflow in function_data is a row named run. Nothing downstream — the tri-comparison function list, the archetype classifier, the call graph, --llm-only output — can tell one CI step from another.

The name is right there

Both major YAML CI dialects name their steps, one line above the run::

- name: Build the wheel        # GitHub Actions
  run: python -m build

- script: python -m build      # Azure Pipelines
  displayName: Build the wheel

GitLab is stronger still: a job's own key is its name, and script: is a child of it.

Candidate shape

Give func_start a capture group fed by the adjacent name key, falling back to the current behaviour when there is none:

"func_start": re.compile(
    r"^[ \t]*-?[ \t]*(?:name:[ \t]*(?P<n>[^\n#]+?)[ \t]*(?:#.*)?\n[ \t]*)?"
    r"(?:run:|script:|before_script:|after_script:)[ \t]*[|>]*",
    re.M,
),

with the displayName:-after-script: (Azure) and job-key (GitLab) forms as the two follow-ups the audit should price. Where no name exists the slice keeps run, so nothing regresses — but then it should also be added to the keyword-bucket exclusion so it is censused consistently with dockerfile's RUN rather than accidentally, by name length. The cleanest form of that is to fix _closed_literal_capture to see through the -?[ \t]* prefix, which is a one-line change and makes the exclusion derived from the rule rather than coincidental.

Strict tests in tests/extraction/languages/test_yaml_strict.py: a named step, an unnamed step, a step whose name: carries a # comment, a name: belonging to the job rather than the step (must not be captured), and a matrix step where name: interpolates ${{ matrix.os }}.

Corpus pairing

keyword-rosetta/data/yaml's steps carry no name: today, so the plant is one name: line per probe — and it needs screening first: name: is not in yaml's doc rule (that one matches description:), but screen_plant.py should confirm it against the full rule set before the shells change. With names, the thirteen slices become distinguishable, the orphan census runs, and raw_state_slop_orphans moves off −100%. Ledger: the yaml half of orphan-detection-is-name-recurrence (shape 2, "SLICE NAME SYNTHETIC OR SHORT") narrows or retires.

Part of #2669 (rosetta[yaml] is #2606). Adjacent to #2728 (keyword buckets) and #2754 (the span-scoped orphan test).