Accept loop variables in the spelling examples actually use (#469)

[jeremymanning]

Closes #469.

The bug

The runtime binds each loop variable under two names — item and $item. The template validator knew only one of them:

self.loop_vars = {'$item', '$index', '$is_first', '$is_last', '$iteration', '$loop'}

So {{ item.name }} inside a for_each step — the form every example in the catalogue writes — was reported as Undefined variable: 'item'.

This is the same false-positive class removed in #448 (slugify), #450 (now), #454 (execution.timestamp) and #459 (pipeline_id): a name the runtime provides, rejected by validation.

The second defect underneath it

Adding the bare names did not make the affected examples pass — it changed their error message. The walker decided a step was a loop with:

is_loop = 'for_each' in obj or 'while' in obj

A step written with foreach — an alias the compiler accepts — was not a loop as far as the validator was concerned, so its loop variables became Loop variable used outside of loop context. Same wrongly-rejected pipeline, different complaint. A validator that knows one spelling and a compiler that accepts another is #465 one layer down.

LOOP_STEP_KEYS now covers for_each, foreach, while, create_parallel_queue, action_loop.

Correcting the issue

Issue #469 as I filed it predicted that accepting bare names would trade a false positive for a false negative. That was wrong. The in_loop_context gate already existed, so {{ item }} outside a loop is still an error, and a pipeline that declares its own input named item still wins over the loop reading. Both are pinned by tests.

One declaration

Both name sets move into core.template_globals alongside the rest of the pipeline language. data_flow_validator imports the same object rather than keeping an equal copy — the test asserts identity, not equality, because an equal copy is exactly what drifted into #469.

Baseline writer bug, found while updating the baseline

--update regenerates the file in full, header included. The header had been extended by hand when examples/supported/ landed, so my first --update silently deleted the paragraph distinguishing "passes validation" from "is a supported example" — an 8-line diff for two added entries.

Hand-fixing the file would have restored the text and left the trap armed. The header moves into BASELINE_HEADER, and a new test asserts the bytes on disk equal what write_baseline() would produce, so any hand-edit to the generated file is a red test rather than a deletion later.

Verification

Three mutations, all caught:

mutation	result
revert to the $-only name set	16 failures
only for_each/while count as loops	3 failures — exactly the three aliases
remove the declared-input guard	1 failure

• Blocking suite: 868 passed, 13 skipped
• Lint: clean (E9,F63,F7,F82,F821,F823,F601,F811)
• Catalogue: 50 → 52 validating; all 52 listed still validate

examples/web_research_pipeline.yaml now fails only on current_date, a name the runtime genuinely does not provide — a real defect in that example, correctly reported.

🤖 Generated with Claude Code