AOT M1-G: one matching rule for every built-in type name

[Aaronontheweb]

Part of the AOT milestone-1 stack (stack #8607). Sits on top of #8608.

Problem

The stack had two rules for matching a HOCON type name against Akka's built-in types:

• Dictionary tables (loggers, scheduler, mailboxes, message-queue semantics, routers, supervisor-strategy configurators, serializers and serialization bindings) required exactly Ns.T or Ns.T, Akka, after StripAssemblyIdentity.
• BufferPools.MatchesBuiltIn, DnsExt and the serialization-identifiers matcher split at the comma and compared the assembly name case-insensitively, so they also accepted Ns.T,Akka and Ns.T, akka, as Type.GetType did.

With Akka.DynamicTypeLoading on nobody could tell, because a table miss falls back to Type.GetType. With it off, "Akka.Dispatch.BoundedMailbox,Akka" threw at a mailbox site but was accepted as a buffer-pool class.

Change

• Two internal helpers in Akka.Util.TypeExtensions:
  • TrySplitTypeName strips the assembly identity, then splits at the first comma outside a generic argument list. The split used to be private to SerializerIdentifierHelper.
  • ToBuiltInAkkaTypeName returns the bare type name when the assembly is absent or is Akka (case-insensitive), and null for any other assembly.
• Every table has one key per type (the bare name), and every lookup goes through ToBuiltInAkkaTypeName. Serialization bindings accept the four framework assembly names instead.
• BufferPools.MatchesBuiltIn is deleted. BufferPools.Create, DnsExt and SerializerIdentifierHelper use the shared helpers.
• The eight copies of the "two spellings are deliberate" comment are gone.
• Net: about 35 fewer lines of product code in src/core/Akka.

With the switch on, nothing changes. With it off, every site accepts the spellings Type.GetType accepted for Akka's own types. Ledger: new G row, and the A row no longer claims two keys per table.

A bare name (no comma) is not trimmed. HOCON keys arrive untrimmed, and trimming would let a padded akka.actor.mailbox.requirements key collide with core's own entry. An existing spec guards this.

Tests

• Exact input → output theories for TrySplitTypeName / ToBuiltInAkkaTypeName: bare name, , Akka, ,Akka, , akka, versioned AQN, generic name with bracketed arguments, another assembly, empty input.
• With the switch off, "Akka.Dispatch.BoundedDequeBasedMailbox,akka" resolves as a mailbox type, and "Akka.Routing.RoundRobinPool,Akka" deploys as a router. Both fail on AOT (7/7): canary in CI with a warning baseline, buildTransitive .targets for automatic AOT/trim opt-in, Native AOT docs #8608.

Verification (local)

• Core builds with -warnaserror at 0/0. The solution builds.
• Akka.Tests: 1425 passed, 0 failed.
• Akka.API.Tests passes unchanged. No public API change.
• AOT canary: unrooted publish runs and prints [canary] OK. Rooted publish shows no new warnings against the 12-entry baseline.

[Aaronontheweb]

Internal type matchers only need one record now - don't need the extra assembly name qualification. Matches what our reflection-based system did before.

LGTM