perf: extend the walker-rule prefilter to call expressions

[mattiamanzati]

What changed

Follow-up to #583. The reference-node prefilter left call expressions ungated — and calls were the largest remaining source of rule-phase GetTypeAtLocation cost, because querying a call's type re-runs expensive call-checking machinery.

• NodeCouldBeStrictEffect now handles KindCallExpression via a new callCouldReturnStrictEffect: it consults the cached resolved signature and its cached return type (both computed during the main check phase, so the lookup is nearly free) and rules the call out when the declared return type conclusively cannot be a strict Effect, using the same conservative couldBeNamed walk as perf: prefilter walker-rule type queries by declared type #583. Both walker rules pick this up through their existing gates with no rule changes.
• Resolved signatures are instantiated, so generic calls are handled precisely: generic("x") conclusively returns the literal "x" and is ruled out, while signature-less calls, any, Effect-containing unions, and optional chains (return type union-widened with undefined) all stay conservative.
• promiseInEffectSuccess also stops computing GetTypeAtLocation for call nodes only to discard it — it goes straight to the resolved signature's return type, falling back to the location query only when no signature resolves.

Soundness

A call expression's type is its resolved signature's return type (promiseInEffectSuccess already relied on this equivalence), calls are not flow-narrowed references, and StrictEffectType requires the type's symbol to be named Effect. The gate answers a conclusive no only when the declared return type provably contains no possibly-Effect constituent; every shaky case falls through to the full query, and the whole path is panic-guarded like the rest of the predicate.

Verification

• Emitted diagnostics (errors and warnings) are byte-identical across a full tsc -b --diagnostics build of the Effect monorepo.
• Measured on that build (3-run medians, matched stock controls): a further −1.11s (−4.6%) on top of perf: prefilter walker-rule type queries by declared type #583. Against a pristine tsgo binary built from the same submodule commit and toolchain, total Effect diagnostics overhead is now +17.0% (19.359s → 22.657s), down from +41.7% before perf: skip rules below the minimum visible severity #581/perf: prefilter walker-rule type queries by declared type #583.
• New unit tests cover the call gate: non-Effect-returning calls ruled out, instantiated generics ruled out precisely, and Effect returns, unions, any, and optional chains kept conservative.
• pnpm lint, pnpm check, and the full pnpm test suite pass.

🤖 Generated with Claude Code