fix: fail loudly on an unknown workflow expression filter

[doquanghuy]

Description

Fixes #3073.

The expression evaluator's filter dispatch fell through to return value for any unregistered filter, so a typo'd or unsupported filter — e.g. {{ items | length }} — rendered the value unchanged with no error, and the run completed. A silent wrong result, with nothing pointing at the cause.

This makes the unknown-filter path fail loudly: an unrecognized filter raises a clear ValueError naming the offending filter and the valid ones, reusing the same strict-ValueError style already used for from_json (which raises on its mis-wired forms precisely to avoid "silently falling through to the unknown-filter path and returning the unparsed value"). It closes that path in general, consistent with the fail-loud direction of #2957 (fan-out items) and #2961 (from_json).

• The five registered filters (default / join / map / contains / from_json) are unchanged — no behavior change for any valid usage.
• Both forms of an unknown filter are caught: the no-arg form (| length) and the name(arg) form (| upper('x')), which previously also fell through silently.

Backward-compat note (calling it out honestly): this turns a previously silent author error into a loud one. Any workflow that today relies on an unknown filter being silently dropped would now error. That passthrough was never intentional (the from_json strictness comment already names it as a path to avoid), and a scan of the repo's own workflow/template YAMLs found no {{ … | <unknown-filter> }} usage, and the full suite stays green — but flagging it so you can weigh the direction. Happy to rework as a warning-instead-of-raise, or gate it, if you'd prefer.

Testing

• Tested locally with uv run specify --help
• Ran existing tests with uv sync && uv run pytest — full suite 3912 passed, 59 skipped
• Tested with a sample project (if applicable) — ran a workflow whose shell step used {{ [1,2,3] | length }}: pre-fix it printed count=[1, 2, 3] and completed; post-fix the run fails loudly with unknown filter 'length'.

Three new tests in TestExpressions (tests/test_workflows.py):

• test_filter_unknown_name_raises — | length raises ValueError.
• test_filter_unknown_name_with_args_raises — | upper('x') raises (the name(arg) form).
• test_registered_filters_unaffected — regression: all five registered filters still work.

The two unknown-filter tests are red against main (DID NOT RAISE), green with the fix.

AI Disclosure

• I did use AI assistance (describe below)

This change was authored with Claude Code: the AI drafted the fix and the tests and ran the suite. I discovered the behavior through end-to-end workflow-expression testing, verified the root cause and the red-against-main / green-with-fix tests myself, and reviewed the diff. The fix deliberately reuses the existing from_json error style rather than inventing a new one.

[doquanghuy]

@mnriem this is the fix for #3073 — fail-loud on unknown expression filters, reusing the from_json ValueError style; the five registered filters are unchanged and the full suite stays green. Happy to adjust the direction (warn instead of raise, or gate it) if you'd prefer.