perf: prefilter walker-rule type queries by declared type

[mattiamanzati]

What changed

effectInFailure and promiseInEffectSuccess walk every node of a source file and call GetTypeAtLocation — a flow-analysis query, the most expensive part of Effect rule execution — just to test whether the node's type is a strict Effect. This PR adds a declared-type prefilter that skips that query for reference nodes which conclusively cannot be an Effect:

• New TypeParser.NodeCouldBeStrictEffect(node) in internal/typeparser/could_be_strict_effect.go: for identifiers and property accesses it resolves the referenced symbol (via the already-cached ReferenceSymbolAtNode) and inspects its declared type with a conservative walk. false is returned only on a conclusive negative: primitives/never, plain object types whose symbol has a different name, and unions of those.
• Both walker rules gain a one-line gate before their GetTypeAtLocation call. Skipped nodes can never match, so the rules' shouldSkip/matched bookkeeping is unaffected; call expressions are never gated.
• The internal walk is generalized over a type-name set (couldBeNamed), so prefilters for other wrapper types (Stream, Layer, …) are one-line wrappers if ever needed.

Why it is sound

StrictEffectType only matches types whose symbol is named Effect. Flow analysis can only refine a reference's declared type — select union constituents, narrow any/unknown, or intersect it (and intersections are synthetic, with no Effect symbol). So a declared type conclusively containing no possibly-Effect constituent can never produce a strict-Effect flow type. Every case where this argument gets shaky stays conservative and falls through to the full query: any/unknown, type parameters/conditionals/indexed accesses, symbol-less types (including the object keyword and evolving types), unions nested beyond depth 4, unresolvable symbols, and all other node kinds.

Example

declare const message: string          // declared type conclusively non-Effect
declare const eff: Effect.Effect<number>

message  // flow-analysis query skipped by the prefilter
eff      // full query runs as before

Verification

• Emitted diagnostics (errors and warnings) are byte-identical to current main across a full tsc -b --diagnostics build of the Effect monorepo.
• Measured on that build, 3-run medians against current main under matched stock controls: total 26.826s → 24.099s (−10.2%), check 26.721s → 23.778s (−11.0%). Composes with the severity skip from perf: skip rules below the minimum visible severity #581 (these rules are warning-severity, so perf: skip rules below the minimum visible severity #581 does not cover them).
• New unit tests exercise the predicate against real checker types (Effect v4): conclusive negatives, all conservative fall-throughs (any/unknown, type parameters, object keyword, Effect-containing unions), call expressions never gated, and nil safety.
• pnpm lint, pnpm check, and the full pnpm test suite pass.

🤖 Generated with Claude Code