core(reply): typed single-shot reply port — deletes Box<dyn Any> erasure (#115)

[joeldsouzax]

What

Rebuilds reply (#115) as the typed, single-shot reply-channel primitive in bombay-core, following the epic #122 rigor contract. Kameo's Box<dyn Any> Reply-trait erasure (~1,048 LOC, ~40 std impls + a derive) is deleted — a typed per-variant reply port erases nothing, so any R: Send + 'static is a reply.

Depends on #113 (error model), now on main.

Design (spec: docs/superpowers/specs/2026-07-05-reply-channel-primitive-design.md, ADR-0002)

• ReplySender<R, E = Infallible> / ReplyReceiver<R, E> / reply_channel() over tokio::sync::oneshot<Result<R, E>> — the oneshot kept an impl detail behind the wrappers (mailbox channel-seam philosophy, ADR-0001).
• send / send_err consume self → a double-reply is a compile error (proved by a compile_fail doctest), strictly stronger than kameo's runtime "second send is a no-op".
• A gone asker is reported as AskerGone (unit signal) — a reply to a vanished asker is un-actionable, and handing the payload back would force a mutation-dead match arm.
• recv maps the oneshot outcome into core(error): rebuild from kameo reference — typed error domains #113's AskError: Ok(Ok r)→Ok(r) · Ok(Err e)→Handler(e) · sender-dropped→Interrupted. Generic over the never-produced M, so core(request): rebuild from kameo reference — ask/tell builders #118 gets a ready-to-return AskError<M, E>.

Re-scope (recorded on #115)

DelegatedReply / ForwardedReply are deferred to #116/#118 — in the reference they are produced only by Context::reply_sender/forward, which need ActorRef + Message; shipping them now would be dead-until-later types testable only for construction (YAGNI, CLAUDE rule 4). They land wired + tested with their machinery.

Tests (TDD, 4 cross-cutting categories first)

8 tests: @bug typed-handler-error probe, Ok-reply sequence, recv-parks-then-send-wakes (reverse ordering — the oneshot waker path the buffered-value tests skip), drop→Interrupted lifecycle (never hangs), send-to-gone-asker defensive, Infallible (tell) roundtrip, 2-thread barrier'd linearizability, and a proptest over all three handler actions; plus the compile_fail doctest.

Mutation: 0 missed (cargo-mutants on reply.rs). No bombay-owned atomics → loom N/A (delegated to tokio oneshot), same as #113.

Benched (benches/reply.rs)

The typed port is ≈1.5× faster than kameo's Box<dyn Any> erasure mechanism — 21.4 µs vs 32.8 µs per 1k reply roundtrips (the box-on-send + downcast-on-recv the redesign removes). Measured, recorded in ADR-0002 — the redesign's justification, substantiated rather than asserted.

Gate

nix flake check green. No README change — the rebuilt spine is not behind the public umbrella yet (same as #113/#114/#133); coverage recorded in docs/testing/coverage-baseline.md.