[#67] Research: reject the absolute-long-offset VarHandle adapter

[peter-lawrey]

Disposition

Closed unmerged. The experiment was successful, but its result rejects the
implementation proposed by this PR.

Scoped result

The prototype produced the same outputs as the existing Unsafe path in 2,730
exercised comparisons. In the retained JDK 11/17/21/25 matrix, adapting the
current absolute-long-offset Access API to byte-array view VarHandles was
slower for XXH3 at every tested size and alignment.

This does not establish that VarHandle is inherently slower. The experiment did
not cover a specialised relative-int-offset path, XxHash, streaming input,
heap/direct ByteBuffer or raw memory. The adapter also initialises and uses
Unsafe infrastructure, so it is not an Unsafe-denial fallback.

Durable evidence

The cleaned branch contains a research-only archive rooted at develop:

• scoped report and limitations;
• corrected reproduction instructions;
• explicitly activated prototype patch;
• complete aggregate table;
• raw JMH JSON and logs for JDK 11, 17, 21 and 25;
• retained JDK 21 reproduction; and
• equivalence/full-test log.

Archive commit:
https://github.com/OpenHFT/Zero-Allocation-Hashing/tree/2339c0db7fdac48b7564e93aeeb732e6bee0af95/research/issue-67

Follow-up

The complete post-Unsafe architecture is tracked in
#124.

Related to #67. This PR does not merge or close the successor issue.

[peter-lawrey]

This PR is being closed unmerged because the experiment's result rejects the
implementation proposed by the PR.

The durable research archive is available at
https://github.com/OpenHFT/Zero-Allocation-Hashing/tree/2339c0db7fdac48b7564e93aeeb732e6bee0af95/research/issue-67.
It contains the
scoped conclusion, corrected reproducibility instructions, prototype patch,
aggregate table, raw JMH JSON/logs and the equivalence-test log.

The evidence supports a narrow conclusion: adapting the current
absolute-long-offset Access API to byte-array view VarHandles is slower for
XXH3 on the tested JDKs and payloads. It does not establish that VarHandle is
inherently slower, and the adapter cannot serve as an Unsafe-denial fallback.

Successor design work is tracked in
#124.