feat: Added support for all Kafka API older versions.

[ShogunPanda]

Signed-off-by: Paolo Insogna paolo@cowtech.it

Fixes #343.

[mcollina]

Reviewed the src side of this branch and pushed fixes in 08b2d49. Full local suite is green except the GSSAPI tests (no KDC image locally) and two schema registry tests that fail identically on the unmodified branch.

Behaviour change worth calling out for the release notes. listGroups and describeGroups now report the group state exactly as Kafka writes it, in Pascal case. Previously the raw value was upper cased, which turned PreparingRebalance into PREPARINGREBALANCE — a value in neither ConsumerGroupStates nor KafkaConsumerGroupStates, despite being cast to ConsumerGroupStateValue. Verified against a local broker:

states observed via admin.listGroups(): [ 'PREPARINGREBALANCE', 'STABLE' ]

Code comparing against 'STABLE' / 'DEAD' / 'EMPTY' needs updating to 'Stable' / 'Dead' / 'Empty'. Documented in docs/admin.md.

Related: the KafkaConsumerGroupStates values this branch added to the listGroups schema were accepted client side but matched nothing on the broker. Brokers compare the states filter case insensitively but not underscore insensitively:

states=["STABLE"]              -> 1 match
states=["Stable"]              -> 1 match
states=["Bogus"]               -> 0 matches
states=["PREPARING_REBALANCE"] -> 0 matches

They are now translated to the wire values before being sent, with a test covering it.

Also fixed

• allowedConfigSources deduplicated — DYNAMIC_TOPIC_CONFIG and the TOPIC_CONFIG alias both map to 1, and a JSON Schema enum requires unique items.
• appendUUID now rejects anything that does not serialize to 16 bytes. UUIDs carry no length prefix and Buffer.from('my-topic', 'hex') is 0 bytes, so the topic-name-as-topic-id backfill added for Metadata v0–v8 would silently desync a whole request if it ever reached a UUID field. The Fetch v12 remapping covers the known path; this makes the invariant enforceable for the others.
• appendVarIntBytes now skips empty buffers, like appendString and appendBytes already do. Appending one corrupts DynamicBuffer's positional reads (readInt8(2) → 0 instead of 3), and record keys and values may legally be empty without being null. Nothing is wrong on the wire today because the write path drains via slice() and crc32c() reads the flattened buffer. Reported upstream as Appending an empty buffer corrupts subsequent positional reads dynamic-buffer#12.
• The Fetch v12 name→id map is built once per request instead of once per request and once per response, keeping the allocation out of the fetch loop's callback.
• writer.context.requestTimeout removed from Produce v3–v6. Nothing reads it — the connection uses its own requestTimeout option — and v7–v11 never set it. The v3/v4 tests asserting it were removed too; the meaningful noResponse assertions stay.
• The two admin fan-outs now aggregate throttleTimeMs the same way, reducing rather than spreading a caller-sized array.

Not fixed, tracked instead

Pinning ApiVersions to v1 costs clientSoftwareName/clientSoftwareVersion for every user, including on modern brokers, so KIP-511 client metrics go blank. The proper fix is negotiating down from v3 on UNSUPPORTED_VERSION, which also needs api-versions-v3.parseResponse to bail on errorCode before parsing the body. I did not implement it because the CI matrix starts at Confluent 7.5 (Kafka 3.5), where the downgrade path never executes — it would ship untested. Filed as #343, and the tradeoff is now documented at the call site.

Left alone

The three fixup commits are still there. Squashing means force-pushing over your commits while you're away, so I left that to you or to a squash merge.

[mcollina]

lgtm