bench: add Share Consumer (KIP-932) + KPipe KEY_ORDERED arms

[eschizoid]

Summary

Adds two runtimes to the competitive ParallelProcessingBenchmark, building on the now-merged key-ordered work (#137):

• Kafka Share Consumer (KIP-932) — one KafkaShareConsumer; each polled batch is fanned to virtual threads with a per-batch barrier before the next poll (implicit ACCEPT = at-least-once). The controlled apples-to-apples rival of KPipe PARALLEL: both unordered, one process, VT fan-out — only the fetch+acknowledge protocol differs.
• KPipe KEY_ORDERED — same single consumer, per-key serial queues. The only arm with an ordering guarantee; present to quantify the per-key ordering tax, NOT as a head-to-head (no other runtime, share included, can order).

Both reuse the existing Testcontainers harness, so every variable except the consumption mechanism is pinned: same broker, seed, workMicros sweep, and @OperationsPerInvocation throughput metric.

Supporting changes:

• Keyed seeding (KEY_CARDINALITY=1000) so KEY_ORDERED is meaningful; the unordered arms ignore keys.
• Broker env for share groups (share rebalance protocol + __share_group_state topic RF) and a per-group share.auto.offset.reset=earliest set via Admin, so the fresh share group reads the pre-seeded records.
• METHODOLOGY.md: the apples-to-apples framing, the ordering-tax caveat, and the single-process (not scale-out) scope.

Published results

Full publishing run committed in this PR: benchmarks/results/2026-05-29.md + raw results-2026-05-29.json. fork=2, warmup=3, iter=5, gc profiler. Intel i7-8850H (6c/12t), JDK 25.0.2, Testcontainers Kafka 4.2.0.

Throughput (records/s; see the results file for ± error bars):

Runtime	Ordering	workMicros=0	workMicros=100	workMicros=1000
Raw KafkaConsumer + VT	none	630,515	573,101	573,443
KPipe (PARALLEL)	none	525,051	581,500	503,612
KPipe (KEY_ORDERED)	per-key	311,854	240,005	89,745
Reactor Kafka	none	309,724	78,953	9,011
Confluent Parallel Consumer	none	111,541	108,518	64,539
Kafka Share Consumer (932)	none	4,839	4,813	4,789

Takeaways (full detail in the results file):

• KPipe PARALLEL tracks the raw-VT ceiling within ~10% and stays flat under blocking work (Loom).
• KEY_ORDERED's 312k → 90k drop is the ordering tax, not a loss — still ahead of Confluent/Reactor at 1000 µs.
• Reactor collapses 310k → 9k under blocking work (core-bound scheduler); Confluent is pool-bound at ~64k.
• Share is protocol-bound and flat at ~4.8k regardless of workMicros. This is a single share member — the bench does not exercise share's scale-out (members > partitions), so read it as a single-member ceiling in early KIP-932, not "share is slow," and not an ordering verdict against KEY_ORDERED (they aren't rivals).

Caveats (also in the results file): Intel x86 not Apple Silicon (don't compare absolute numbers across machines); workMicros=0 is the noisiest cell; macOS + Docker loopback broker means absolute records/s are low — read orderings, not absolutes.

Test plan

• Both new arms validated end-to-end on Docker (share group forms, assigns all 8 partitions, consumes seeded records).
• Full publishing run + dated results file committed.
• Follow-up (separate): multi-member share-group experiment to evaluate share scale-out (its actual value prop).
• Follow-up (separate): tune share throughput (max.poll.records, ack batching, acquisition-lock timeout).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (main@54bf55f). Learn more about missing BASE report.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #138   +/-   ##
=======================================
  Coverage        ?   74.64%           
  Complexity      ?      763           
=======================================
  Files           ?       76           
  Lines           ?     2930           
  Branches        ?      367           
=======================================
  Hits            ?     2187           
  Misses          ?      568           
  Partials        ?      175           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.