check_interface()'s token matching reads YAML block-scalar content as structure #949

Found on PR #948 (CodeRabbit review) while adding a new requireTokensInJob/forbidTokensInJob entry for the deploy-site caller-stub interface contract, but the weakness is in check_interface()/_code_view() generally, not specific to that one entry.

The gap

_code_view() (spec/audit.py) strips comment-only lines (#-prefixed), but does not strip YAML block-scalar (|, >, and their chomp-indicator variants) body content. Every requireTokensInJob/forbidTokensInJob check in check_interface() is a plain substring search over that view, so a job whose name: (or any other string-valued key) is written as a block scalar containing text that happens to match a required or forbidden token, at a matching indent, can satisfy or trip the check without the actual YAML structure the token is meant to verify existing at all. Concretely, for the deploy-site.yml contract's new "with:\n environment:" token:

  deploy:
    name: |
      with:
      environment:

check_interface() reports this caller stub as carrying the required with.environment input. It does not - the job has no with: mapping at all, and would fail at GitHub Actions dispatch time with a missing-required-input error, exactly the failure mode #942's fix was meant to catch mechanically rather than only at runtime.

Scope

This is not unique to the deploy-site contract. Every requireTokensInJob/forbidTokensInJob entry across spec/files.json reads through the same _code_view() substring search, so the same construction (a block scalar crafted to contain a matching token at a matching indent) can fool any of them, not only the newly-added one.

Ask

Decide the proportionate fix. Two shapes, roughly by cost:

1. Targeted mitigation: strip YAML block-scalar body lines (a key: |/key: > line followed by more-indented lines, until dedent) from _code_view()'s output, the same bounded-scope heuristic the function already uses for comments. Catches the realistic case without a general YAML parser.
2. Structural validation: parse the job mapping properly (e.g. via a YAML library, scoped to just the job block) and check with.environment as an actual child key rather than by substring, which CodeRabbit's own review labeled a "heavy lift" - likely the right call only if _code_view()'s scope of false-positives/negatives turns out to be broader than block scalars alone.

Given every interface check across the file shares this weakness and the trigger requires a deliberately adversarial construction (not something ordinary caller-stub authoring produces by accident), I judged a full engine rewrite out of scope for #948's actual bug fix and filed this separately rather than block that PR on it.