Fix completion value clobbered by re-entrant Evaluate() during module execution

[lahma]

Summary

Fixes #2492 — a regression where a script's completion value could be clobbered by a re-entrant Evaluate() triggered from module execution, so a previous evaluation returned the wrong object (reported as Cannot read property 'generateRandomText' of undefined).

Root cause

Engine.ScriptEvaluation stored the script's completion value in the engine-level _completionValue field and then drained the shared event loop:

_completionValue = result.GetValueOrDefault();
RunAvailableContinuations();   // can re-enter the engine
return this;                   // caller then reads engine._completionValue

RunAvailableContinuations() can re-enter the engine: an awaited dynamic import() runs a module whose body calls a CLR function that itself calls Evaluate(). Because the event loop has an _isProcessing re-entrancy guard, the middle Evaluate() is the active drainer; the deeper Evaluate() writes _completionValue before the middle frame reads it back — so the middle Evaluate() returns the deeper frame's value. (Surfaces only when the imported module has its own import, which shifts microtask timing.)

Fix

Carry the completion value out as a per-frame local instead of a shared field — a local is immune to re-entrancy by construction:

• ScriptEvaluation now returns the JsValue (captured before the drain, returned after).
• Evaluate(in Prepared<Script>) and Execute(in Prepared<Script>) funnel through a shared ExecuteForCompletion(...) helper.
• ExecuteAsync(string,...) reads via Evaluate(...).
• The _completionValue field is removed.
• The invariant for the remaining ambient fields (_error, _lastSyntaxElement) is documented; they are safe because they're produced and consumed synchronously with no re-entrant drain in between.

No public API signatures change.

Tests

Two regression tests, both verified red on main, green after:

• EngineTests.NestedEvaluateDuringContinuationDrainDoesNotClobberOuterResult — minimal deterministic repro (Promise.resolve().then queues a microtask drained by a nested Evaluate()); no modules needed.
• ModuleTests.NestedEvaluateFromModuleExecutionDoesNotClobberCompletionValue — faithful repro of the issue (dynamic import + import * as + an importLibrary CLR callback).

Full Jint.Tests passes (3030); Test262 unchanged (99,206 passed; only the pre-existing timing-flaky Atomics.waitAsync differs, confirmed flaky on main too).

Performance

The change removes a field rather than growing any hot struct (e.g. ExecutionContext, which is full-copied on every push / scope change), so no regression is expected. MinimalScriptBenchmark (BenchmarkDotNet ShortRun, net10.0), base vs fix:

Method	Base Mean	Fix Mean	Base Alloc	Fix Alloc
Execute	2.892 us	2.917 us	17.24 KB	17.23 KB
Execute_ParsedScript	1.792 us	1.766 us	15.33 KB	15.32 KB

Means are within the measurement error; allocations are flat-to-marginally-lower.

🤖 Generated with Claude Code