fix(typeselector): drop a generic candidate the field cannot close

[VPDPersonal]

Summary

The candidate scan matched the two generic definitions and ignored their arguments, so a candidate that fixes an argument itself was offered for a field it can never fit:

	Before	After
IConverter<float, float> field	GenericToString<TFrom> listed	dropped
After picking it	argument page refuses every choice	—

GenericToString<TFrom> : IConverter<TFrom, string> is an IConverter<,>, yet no TFrom makes the closed type assignable. Inference then failed (correctly), the caller fell back to the open definition, and the row became a dead end. The whole numeric converter family in Aspid.MVVM showed it.

• 🎯 CanCloseToFieldType now compares the arguments too, position by position, rather than matching definitions alone. GenericBaseDefinitions became dead and is removed.
• ↔️ The comparison honours declared variance, because assignability does. With IConverter<in TFrom, out TTo> an IConverter<object, string> candidate is still listed for an IConverter<string, string> field. A purely structural comparison would have traded this defect for its mirror image — a usable candidate missing from the list, which is not hypothetical: an IConverter<float, object> field legitimately accepts GenericToString<float>.
• ⛔ It also honours where variance stops. The CLR applies variance only across an implicit reference conversion, so a position the field closed over a value type admits exactly one argument, like an invariant one. Note that IsAssignableFrom alone does not express this — typeof(object).IsAssignableFrom(typeof(int)) is true by boxing — hence an explicit guard.
• 🔁 Pinned positions are matched first and their bindings recorded; variant positions are judged afterwards, against what the pinned ones already forced. That ordering is what also drops SequenceConverters<T> : IConverter<T, T> from an IConverter<float, string> field — the same dead row, found while verifying this fix — without depending on which position happens to be declared first.
• ↩️ A position that still admits a family of arguments is left alone. Pair<TKey, TValue> under an IKeyed<string> field keeps its open definition and still opens the argument page. Proving no argument converts would mean sweeping the domain — work the page already does, validating each choice through TryConstruct.

Verification

Live 6000.4.0f1 Editor, against Aspid.MVVM's real IConverter hierarchy:

Field	Candidates after the fix
IConverter<float, float>	GenericFuncConverter<Single, Single>, SequenceConverters<Single> — GenericToString gone (the reported defect)
IConverter<float, string>	GenericFuncConverter<Single, String>, GenericToString<Single> — SequenceConverters<open> gone (side finding)
IConverter<float, object>	GenericFuncConverter<Single, Object>, GenericToString<open> — variance kept the usable row
IConverter<string, string>	unchanged, all three closed

Check	Result
EditMode suite	✅ 352/352
New cases fail against the previous behaviour	✅ 4/4 negative cases
Brute-force cross-check over 7 field shapes (every candidate closed over an argument pool, tested with IsAssignableFrom)	✅ no over-rejection

Nine cases are added to GenericTypeResolverTests, covering the repro, the inverse (candidate fixes an argument the field agrees with), the determined and partially-determined candidates, variance on a reference type, two value-type rejections, and a definition implemented twice.

Notes for review

• ✅ run_tests can be driven from this CLI, contrary to the note in fix(typeselector): list a determined generic candidate closed #173: run_tests --mode EditMode --async_tests true plus polling test_status works. Only the synchronous mode hits the Pipeline server's 30 s cut-off.
• ⚠️ Not exercised through the picker UI. Verification went through direct resolver calls on a live Editor, not by clicking the dropdown. A manual check is added to the QA checklist.
• 🔎 One residual class of dead rows is left deliberately. When a single unconstrained parameter sits on two variant positions (SequenceConverters<T> under an IConverter<object, string> field), the pre-filter still passes it — the brute-force sweep reports offered=3 against closable=2. Ruling it out means enumerating the domain, which is the argument page's job; the behaviour predates this change and is documented in the <remarks>.
• 📋 Docs synced per repo rules: CHANGELOG EN/RU, QA-CHECKLIST EN/RU, SerializeReferences.md EN/RU.

🤖 Generated with Claude Code