okhttp: avoid unnecessary Header allocation when indexing already-lowercase names in Hpack.Writer

[jaipilot]

What

Builds on grpc#12907 ("okhttp: optimize HPACK to index :path and fix dynamic table eviction bugs"), which fixed a bug where headers were inserted into the HPACK dynamic table using their original-case name instead of the lowercased name used for indexed lookups, by unconditionally allocating new Header(name, value) at both dynamic-table-insertion call sites in Writer.writeHeaders.

This change avoids that allocation in the common case: ByteString.toAsciiLowercase() returns this (same reference) when the name has no uppercase ASCII bytes, which is true for gRPC's generated header names and the newly-indexable :path/:authority pseudo headers. A small canonicalHeader(header, name, value) helper reuses the original Header instance when name == header.name, and otherwise allocates exactly as before (preserving the original PR's case-folding fix for mixed-case names).

Why

Header.equals/hashCode/hpackSize are purely value-based (over name/value content), and Header is immutable, so reusing the original instance when its name is already the indexing name is behavior-preserving: identical bytes are written, identical dynamic table state (size, order, byte accounting) results, and identical case-insensitive matching semantics apply.

Evidence

• Behavior lock: added HpackTest#mixedCaseHeaderNameIndexedOnRepeat, verifying a mixed-case header is stored lowercase and recognized as an indexed reference on repeat. Passes identically before and after the change (59/59 HpackTest tests both times).
• Performance: an instrumented allocation counter showed writeHeaders allocating 5 Header objects (one per header) on the original code for a representative gRPC header block, versus 0 on the optimized code, reproduced across multiple runs. A ThreadMXBean allocated-bytes benchmark (5 runs x 200k iterations each side) showed a consistent 8488 -> 8368 bytes/iteration reduction.
• Full grpc-okhttp module build (compile, test, checkstyle) passes: 360 tests, 0 failures/errors.

Limitations

• The performance benchmark was a standalone, non-committed microbenchmark (not part of this diff); only the deterministic allocation-count result and its methodology are reported here.
• No change to public API, wire format, or dynamic-table semantics; scope limited to the two Writer.writeHeaders call sites and the eviction/indexing logic already touched by the evaluated PR.

---

Generated by JAIPilot Cloud for #1 from Anthropic session sesn_01758uGvBonuiwxT2TZEkHEP.