Compile event-loop metrics, predicate-narrowed catch bindings and other wrapper-reachable gaps

[cramforce]

Stacked on #1413 (variadic generic wrappers). Once wrapper bodies compile, the code behind them becomes reachable, and with it a set of independent gaps. This PR closes the general ones.

Summary

• performance.eventLoopUtilization() (no-argument, one-argument and two-argument difference forms) and monitorEventLoopDelay() (an IntervalHistogram with enable, disable, reset, percentile, and the min, max, mean, stddev, count and exceeds readings). These work under the shipped fallback declarations and under a project's own @types/node.
  • The histogram is a numeric handle into a new runtime unit, scr_perf_loop.c. The unit samples on an unref'd interval at the requested resolution and records each interval in nanoseconds, like Node's ELDHistogram. Values land in HdrHistogram-shaped buckets (three significant figures), so the readings are the bucket values Node reports. An empty histogram reads min as 2^63−1 and mean/stddev as NaN, like Node. percentile validates Node's range and throws ERR_OUT_OF_RANGE.
  • Utilization follows Node's algorithm: zeros before the loop starts, and the one- and two-argument difference forms. Active time is the main thread's CPU time since the loop started, which differs from Node's poll-phase accounting (documented as a limitation). The emitted main marks the loop's start only in programs that read the metric.
  • Linking: the unit links only for programs that use these APIs, under a new runtime feature, perfLoop.
• Catch bindings narrowed by a user type predicate or an assertion (if (isErrorLike(e)) e.message, TypeError & { code: unknown }) read their members through the checked-dynamic value, as JavaScript reads them. Before, these reads were SC1063.
• ?? with a function default of a different signature (opts.tags ?? (() => ({}))). The result is the callback's own function type, and the default adapts into it. Before, the checker's merged signature caused SC1090 and would have lost the callback's result shape.
• Optional calls on interface-typed receivers (tracer.scope().active()?.addTags(tags), s?.m(x)) dispatch like the non-optional protocol call. Before, these were SC1090.
• any in TypeScript function types: a parameter typed any (optional or not) of a function declaration or a spelled function type is the checked-dynamic slot, as an any binding already is. A generic instantiated at any (for example a type argument inferred from an untyped forwarding call such as tracer.trace(name, () => fn.apply(this, args))) takes unknown for that parameter or result.
• 'key' in err on Error receivers for keys other than code, message and name. The answer is true when the static class declares the key, otherwise a runtime class test over the subclasses that declare it.
• Object.entries over a generic record parameter: keyof T extends never ? unknown : T[keyof T] resolves through the binding. Fields of different scalar types map to their union.

Forms not covered keep their refusals:

• Pick<SomeClass, "method"> record members still don't accept class instances. Mapping them to the class would refuse object-literal values that compile today.
• Overriding a method whose base returns any (or otherwise differs in its native ABI).

Evidence

• New corpus programs (Node vs native, byte-identical, plain and sanitized lanes): 6390 (predicate-narrowed catch bindings), 6391 (?? function defaults), 6392 (optional protocol calls), 6393 (any parameters), 6394 (in on Error subclasses), 6395 (generic Object.entries), 6396 (generic method at any), 6397 (event-loop metrics).
• @types/node fixtures event-loop-metrics.ts and timeout-type.ts, run by tests/harness/project-config.test.ts.
• Zero-cost gate against Specialize variadic generic wrappers to the wrapped function's signature #1413 (scripts/zero-cost-gate.mjs, bench + self-hosted compiler + all corpus entries, sharded over three sandboxes):
  • IR, LLVM and release executables are identical for every pre-existing program. The 8 new corpus programs are refused by the baseline.
  • Compile time: bench geomean −0.07% (95% CI −0.76% to +0.68%), self-hosted compiler +0.57%.
  • An intermediate run caught three issues, each fixed and re-checked:
    • The new runtime unit added a file symbol to programs that link dynAsync. It is now its own perfLoop feature, linked only where used.
    • The T[keyof T] union widened a literal-key read in 2550-generics-keyof-pick.
    • The any parameter rule changed two object literals of any-taking lambdas and methods (checked-callback-options, 5197); it now covers only function declarations and spelled function types.
• pnpm manifest, pnpm node-compat, pnpm node-compat:check, lint, format, pnpm test:ts7 --baselines-only, diagnostics, frontend unit tests, and the self-hosted IR build all pass. The compatibility diff marks performance.eventLoopUtilization, monitorEventLoopDelay and four Histogram methods partial, with the new tests as evidence.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
scriptc	Ready	Preview, v0	Oct 20, 2026 7:08am UTC