refactor(executor/core): dyn-safe RuntimeOps + de-erased builder & registrar lifetimes (034 Phase 2, #130)

[lxsaah]

Closes #130. Part of design 034 Phase 2 (review doc §3.2/§3.5); first of three stacked PRs (#130 → #133 → #134).

What

1. Dyn-safe RuntimeOps capability trait (aimdb-executor) — groundwork for #131, not yet consumed by core.

• name() / now_nanos() (monotonic, arbitrary epoch) / unix_time() / boxed sleep(core::time::Duration) / log(LogLevel, &str).
• Implemented for TokioAdapter (OnceLock<Instant> anchor), EmbassyAdapter (boot-anchored uptime, µs granularity), WasmAdapter (Performance.now(), wall clock from Date.now()).
• Per-adapter impls are intentional: a blanket impl over TimeOps is impossible (now_nanos needs an epoch anchor; TimeOps::Instant is opaque).
• Shared behavioral contract test (aimdb_executor::test_support) run by all three adapters under their own executors. The embassy host time-driver stub now wakes immediately on schedule_wake so Duration::ZERO sleeps complete (host clock is pinned at 0).
• BoxFuture's canonical definition moves to aimdb-executor; aimdb-core re-exports it (same type).

2. De-erased builder internals (aimdb-core/src/builder.rs) — no behavior change.

• spawn_fns/start_fns stored as Vec<(StringKey, SpawnFnType<R>)> / Vec<StartFnType<R>> instead of Box<dyn Any + Send>; both panicking downcasts in build() deleted.
• AimDb<R>'s struct-level R: RuntimeAdapter + 'static bound moved to its impls (strictly more permissive) so the field types are well-formed on the NoRuntime typestate, where both vectors are provably empty.

3. Registrar lifetime fix (aimdb-core/src/typed_api.rs) — breaking for code that names the old signatures; closure-based configuration compiles unchanged.

• All fluent methods: &'a mut self -> &'a mut Self → &mut self -> &mut Self. A configure closure can now write reg.source_raw(…); reg.tap_raw(…); as separate statements and reuse the registrar after a chain.
• configure's bound drops the HRTB: FnOnce(&mut RecordRegistrar<'_, T, R>).
• OutboundConnectorBuilder/InboundConnectorBuilder: double-'a split into <'r, 'a, T, R>; finish() returns &'r mut RecordRegistrar<'a, T, R> (chains still work).
• RecordT::register, impl_record_registrar_ext!, EmbassyRecordRegistrarExtCustom, and RecordRegistrarPersistExt follow.

Acceptance criteria (from #130)

• Arc<dyn RuntimeOps> constructible from all three adapters; shared behavioral test
• No Box<dyn Any> fields in AimDbBuilder; no downcasts in build()
• reg.source_raw(…); reg.tap_raw(…); as separate statements (new registrar_allows_separate_statements test; multiple_link_from_allowed extended)
• No behavior change: make check green (fmt, clippy matrix, full test matrix, embedded thumbv7em, wasm, deny)

Notes for reviewers

• The ext_macros.rs dissolution into plain generic impls stays deferred to refactor(core)!: remove runtime type parameter R from the object graph — non-generic AimDb/TypedRecord/RuntimeContext (R-removal part 2) #131 (part 2), per the issue.
• Pre-existing, unrelated: aimdb-wasm-adapter/tests/transform_join_integration_tests.rs does not compile for native targets (calls .subscribe()/.produce() on the (AimDb, AimDbRunner) tuple from build()); it predates this branch and is outside make check's legs.

🤖 Generated with Claude Code