Return the pooled buffer when encoding fails

[shin4141]

Summary

AbstractEdgeEncoder.useAsByteArray returned a borrowed EdgeBuffer only after the encoding callback and byte copy succeeded. If the callback threw, that buffer was permanently removed from the shared pool. With a one-buffer pool, a later valid encoding then had no buffer to borrow.

Return the buffer in finally. The original exception still reaches the caller, and the next use resets the returned buffer before writing. This change does not alter successful encodings.

Regression

The new test uses a one-buffer pool, writes one value, then asks the encoder to serialize an unsupported object. On main at 0cd8e45a4bd8f383bf58175e05fe51ba64f0c06b, it fails because the pool size is 0 instead of 1. With the patch, it checks that the original exception is preserved, the same buffer returns to the pool, and a subsequent valid encoding matches a fresh unpooled encoder.

Tests

• ./gradlew :codec-java:test — 67 passed.
• :core:test — 8,918 passed, 5 skipped; :core-java:test — 313 passed.
• ./gradlew :server:test — 745 passed with a test-runtime Byte Buddy agent because the isolated macOS build disallows self-attachment.
• ./gradlew spotlessCheck — passed.
• ./gradlew test — not fully verified: the unrelated engine test fixture's BlockHound.install() cannot self-attach in the isolated macOS environment (Operation not permitted). The five DefaultBufferCapacityTest cases themselves passed.

AI Assistance

• This change was written largely with AI assistance.
  • Tool / model: Codex (GPT-6).

[CLAassistant]

All committers have signed the CLA.

[em3s]

Thanks @shin4141 — the one-buffer test pins it exactly. Fails on main, passes with the patch.

I'll tweak the title and merge. One thing left over: borrow() is pool.poll(), so past poolSize concurrent encodes it returns null and reset() NPEs. Falling back to new EdgeBuffer() would cover it. Happy to take a follow-up PR if you're interested.