Detailed share-group offset mutations

Use AlterShareGroupOffsetsDetailedAsync and DeleteShareGroupOffsetsDetailedAsync when a mixed broker response must retain every confirmed outcome. Both return an AdminMutationResult for every requested entity. They implement the optional IDetailedShareGroupOffsetAdminClient capability; extensions expose them through IAdminClient without adding members to that interface. Custom implementations without the capability throw NotSupportedException.

Alteration results use TopicPartition keys. Alteration requires an empty group and can create an empty group when none exists. Deletion results use ordinal topic-name keys: Kafka deletes all share offsets for a requested topic and returns a topic-level result. It does not report individual partition deletion outcomes. Neither operation is transactional across entities.

using Dekaf;
using Dekaf.Admin;
using Dekaf.Protocol;

ShareGroupOffsetAlteration[] requested =
[
    new() { TopicPartition = new("orders", 0), StartOffset = 42 },
    new() { TopicPartition = new("orders", 1), StartOffset = 84 }
];
var outcomes = await admin.AlterShareGroupOffsetsDetailedAsync(
    "workers", requested, new ShareGroupOffsetMutationOptions { TimeoutMs = 10_000 });

// Only confirmed coordinator rejections are safe to select for this retry.
// The built-in client already performs bounded retries for these rejections.
var retry = requested.Where(item =>
    outcomes[item.TopicPartition].Outcome == AdminMutationOutcome.Failed &&
    outcomes[item.TopicPartition].ErrorCode is ErrorCode.NotCoordinator
        or ErrorCode.CoordinatorNotAvailable or ErrorCode.CoordinatorLoadInProgress).ToArray();
if (retry.Length != 0)
    await admin.AlterShareGroupOffsetsDetailedAsync("workers", retry);

// This removes offsets for every partition of orders, not a selected partition.
var deleted = await admin.DeleteShareGroupOffsetsDetailedAsync("workers", ["orders"]);

Top-level group errors are copied, with their original code and message, to every entity in that request. Confirmed successes and terminal errors survive retries of other entities. Each retry rediscovers the group coordinator. Missing or duplicate response entries are Unknown; unrequested entries are not added to the result dictionary.

A transport failure after entering the send path is Unknown, including cancellation or deadline expiry during the send. It is never automatically replayed. Inspect the group's state before deciding whether to retry an unknown mutation. Cancellation before invocation throws; cancellation during execution returns retained outcomes and NotAttempted results for entities that were never sent. TimeoutMs bounds discovery, sends and retries together; zero expires before discovery. Empty input performs no network operation.

Existing convenience methods retain their existing exception and retry behavior. Existing member-removal, feature-update and Streams-offset result APIs are unchanged.

Dekaf.Testing implements the same result shapes, validates inputs before applying mutations, and retains sibling successes when group/topic/partition fault scopes fail. Active groups reject these mutations; a missing group can be created by alteration, while deletion reports GroupIdNotFound. Deletion keeps the group after removing its final offset. The simulator knows whether cancellation precedes its synchronous state change and reports NotAttempted in that case. Real ambiguous network sends are covered by protocol-response unit tests.

Protocol sources: alteration response, deletion response.