Make Function.prototype.toString() source text retention opt-in (#2560)

[lahma]

Fixes #2560.

Problem

Since v4.9.0 (#2394), Function.prototype.toString() returns the real source text of user functions. To support this, every parsed function node pins the entire source string (Node.UserData = ctx.Input), released only lazily on the first toString() call. Applications that cache many large prepared scripts and never call toString() therefore retain the full source for the lifetime of the cache — the reporter observed ~250 MB of duplicated source strings (the same ~500 KB string retained once per engine).

Fix

Add an opt-in RetainFunctionSourceText option, default false:

• IParsingOptions.RetainFunctionSourceText on ScriptParsingOptions / ModuleParsingOptions — controls per-parse retention (prepared and non-prepared).
• Options.RetainFunctionSourceText (+ fluent options.RetainFunctionSourceText()) — engine-level toggle for the default parser, so bare engine.Execute(string) can opt in without passing parsing options.

When retention is disabled (the default), source text is not kept and toString() returns the pre-4.9 function name() { [native code] } placeholder. When enabled, real source is returned exactly as before.

// lean by default now — no source retained:
new Engine().Evaluate("function f(){} f.toString()"); // "function f() { [native code] }"

// opt back in:
new Engine(o => o.RetainFunctionSourceText()).Evaluate("function f(){} f.toString()"); // real source
Engine.PrepareScript(code, options: new ScriptPreparationOptions {
    ParsingOptions = new ScriptParsingOptions { RetainFunctionSourceText = true } });

Behavior change: this reverts the v4.9 default (real-source toString()) to opt-in, trading it for lower memory by default. The Test262 harness and the existing toString source tests opt in so conformance is unchanged.

Verification

• Jint.Tests: 3147 passed / 0 failed (excluding the pre-existing, environment-sensitive EngineLimitTests.ShouldAllowReasonableCallStackDepth, which crashes on unmodified main here too).
• Test262: 99260 passed / 0 failed — the built-ins/Function/prototype/toString/* conformance tests stay green via the harness opt-in.
• Retained memory (GarbageCollectionTests.PreparedScriptsDoNotRetainSourceTextByDefault): holding 25 distinct ~800 KB cached prepared scripts retains 20.1 MB with retention on vs 1.8 MB off — ~18.3 MB / 91% freed.

The secondary env-pooling suspect (#2413) is left for a separate, benchmarked PR; it is a bounded, self-healing-under-load retention, unlike the quantified source-string regression fixed here.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Bf1LESD4X8dMNPwLKJCAh9

[scgm0]

Can't we just keep the source text once, have the node store its own slice position, and only create the string slice from the source text based on the position when calling toString()?

[lahma]

You can always propose a PR, but sounds that we are still pinning the source string.