SDK DX: anyhow interop, bare check!, macro error spans, suite-level tags

[lann]

From the adversarial review (2026-08-04). SDK/macro ergonomics found by writing probe suites; all verified against real compiler output. Ordered by expected M1 porting pain:

1. ? does not work with anyhow::Error in case bodies: the blanket impl<E: std::error::Error> From<E> for Failure (crates/component-test-core/src/verdict.rs) excludes anyhow::Error (which deliberately doesn't implement Error), and the workaround type Failure isn't in the prelude the macro injects. Porting a conformance suite means fallible helpers everywhere; today every call site needs a hand-rolled map_err plus tribal knowledge of the import. Proposal: prelude extension trait (or_fail() on Result<T, E: Display>) and/or export Failure in the prelude; document the anyhow gap on failed().
2. check!(cond) without a message dies with unexpected end of macro invocation — everyone arriving from assert! writes this first. Add a ($cond:expr $(,)?) arm using stringify!($cond) as detail, or a targeted compile_error!.
3. Diagnostic-anchor spans: the decline-pair lint anchors at Span::call_site() (attribute line) instead of the first offending tagged case (spans are already collected); verdict-shape/type errors also land the primary caret on the attribute (arity is checked syntactically, types aren't) — in a 50-case file users read secondary notes to find the case. Anchor at the case fn (e.g. a typed local let __f: fn(&TestContext) -> _ = #path; spanned at the fn).
4. No way to tag a whole suite: root-level #[cases(tags(...))] dies with cannot find attribute. Support #[suite(tags(...))] or emit a targeted "put it on an inner module" error.
5. Non-leaf WIT-label violations point at the case fn, not the offending module ident; and NameError::BadChar says contains forbidden character 'A' without stating the grammar — the message isn't frozen, improve it.
6. The two example suites model two diagnostic APIs: fixture-suite (the #[suite] DX showcase) uses raw ctx.diagnostic("...".into()) while sample uses the curated ctx.diag("..."). Unify on diag (goldens pin output, not source — no regeneration needed).

Related: #28.

[lann]

Items 1, 2, and 6 landed in 38faf9a on adversarial-multi-agent:

• or_fail() extension trait + Failure in the prelude ({:#} flattens anyhow chains onto one line; contract pinned by test against anyhow itself)
• bare check!(cond) with stringified-condition detail
• fixture-suite unified on ctx.diag() (goldens confirmed output-neutral)

Remaining in this issue: item 3 (diagnostic-anchor spans for the decline-pair lint and verdict-shape errors), item 4 (suite-level tags), item 5 (module-ident spans for non-leaf violations + the BadChar message wording).

[lann]

Span work landed (450d94d, eab60b0): the decline-pair lint anchors at the first case carrying the unpaired positive tag (fell out of routing the lint through core::tags::unpaired_positive_features), and non-leaf WIT-label violations anchor at the offending module ident (mod x__y gets the caret, not a case fn below it). NameError::BadChar now states the segment grammar. All pinned by trybuild (.stderr diffs document the moves; 13 UI cases).

Remaining here:

• item 3's second half: verdict-shape type errors still anchor at the attribute (needs quote_spanned surgery in the registration glue, or a syntactic param-type check)
• item 4: suite-level tags (design input on syntax)