Play Framework

perf: resolve optional compression codecs once instead of per pipeline

PF-173: buildHttpContentCompressor probed for the optional codecs by calling
StandardCompressionOptions.brotli()/.zstd() and catching the failure, on every
pipeline construction — that is once per accepted connection, for the life of
the JVM. ZstdOptions' static initializer fails when zstd-jni is absent, the JVM
marks the class erroneous, and every later touch re-throws a fresh
NoClassDefFoundError with a full stack capture, so the cost never tapers off
after warm-up. JFR on a JClaw run counted the throws tracking connections
rather than requests, which is exactly what a per-pipeline probe predicts.

Resolution moves into a lazily-initialized CompressionOptionsHolder, so the
array is built once and every compressor is handed the same instance. Held in a
nested class rather than a static field so it stays lazy: the compressor is not
in the default play.netty.pipeline, and eager init would load brotli4j's native
library at boot on deployments that never compress.

The probe itself is replaced with Netty's own Brotli/Zstd.isAvailable(), which
are cached booleans computed in their own static initializers and never touch
BrotliOptions/ZstdOptions — the classes whose initializers throw. Caching alone
would still pay the throw once per codec and leave the erroneous-class landmine
armed; this way there is no NoClassDefFoundError at all. Both probes are guarded
so a brotli-less deployment does not inherit the old behaviour. The per-codec
catch stays as a net for a classpath that resolves the loader but fails to
construct, and guarantees the holder's initializer cannot throw — an
ExceptionInInitializerError there would be cached forever and recreate this bug
class permanently.

The compressor stays per-pipeline: it extends HttpContentEncoder, which keeps a
per-channel ChannelHandlerContext and encoder queue and is not @Sharable, so the
options array was the only hoistable per-call work. Sharing it is safe — Netty's
ctor reads the array into per-type final fields and neither retains nor mutates
it, and the options objects hold only final primitives.

Adds HttpContentCompressorCacheTest. The bug is functionally invisible
(compression still falls back to gzip/deflate correctly), so the tests assert on
object identity across two separately built compressors, via the real public
entry point. Verified red against the pre-PF-173 body. Optional-codec
expectations are derived from Netty's availability flags rather than hardcoded,
so the suite holds on either classpath.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ma8qUZG39DAkht2oSA7am2