fix(eve): prioritize current hook deliveries over stale backlog

[iroiro147]

Summary

• split session delivery backlog into turn-owned, current-hook, and replaced-hook queues
• classify ready hook reads before selecting the next delivery, so current live input runs before older replaced-hook backlog
• preserve turn-control cancellation/rebuffer ordering and keep the active hook current when a candidate rekey claim fails

Refs #547.

Tests

• pnpm --filter eve exec vitest run --config vitest.unit.config.ts src/execution/session-delivery-hook.test.ts src/execution/workflow-entry.test.ts src/execution/turn-control-receiver.test.ts src/execution/turn-dispatch.test.ts
• pnpm --filter eve exec vitest run --config vitest.integration.config.ts src/execution/session-delivery-hook.integration.test.ts
• pnpm --filter eve typecheck
• pnpm changeset status --since origin/main
• git diff --check

[vercel]

@iroiro147 is attempting to deploy a commit to the Vercel Team on Vercel.

A member of the Team first needs to authorize it.

[vercel]

Suggested change

	await this.next();
	const result = await this.next();
	// `next()` memoizes a single in-flight read in `offered`, and the hook is
	// shared across turns, so several `bufferNext` callers (e.g. an abandoned
	// request from a prior turn plus the current turn) can await the same
	// resolved read. The first caller consumes and buffers it, clearing
	// `offeredRead`; later callers must not consume again (that would throw
	// "Cannot consume a public delivery before it resolves."). They observe
	// the identical resolved result and report the delivery already buffered.
	if (offeredRead === undefined) {
	return result.done ? "closed" : "buffered";
	}

Two bufferNext() awaiters sharing the memoized offered read both call consumeOffered(), so the second throws "Cannot consume a public delivery before it resolves." and crashes the driver loop.