ci: add centralized docs drift check - #669

…e#448)

* Remove per-batch StreamWriter allocation in binary-insert serializers

The query line preceding each batch's rows (e.g. INSERT INTO ... FORMAT
RowBinary) was written via a StreamWriter, which allocated the writer plus a
4 KiB char buffer and a matching byte buffer on every batch. Encode it into a
pooled scratch buffer (new internal QueryLineWriter) and write directly instead,
removing one allocation per batch on both the object[] and POCO insert paths.

As a side effect the query line is now emitted as plain UTF-8 with a '\n'
terminator: the BOM that StreamWriter prepended is gone, and the newline is
'\n' on all platforms (was '\r\n' on Windows via Environment.NewLine).
ClickHouse accepts both, so behavior is unchanged.

Benchmark (200k-row POCO RowBinary insert, uncompressed, ENGINE Null):
  BatchSize 500  (400 batches): 7783.7 KB -> 4503.4 KB (-42%), Gen0 eliminated
  BatchSize 5000 (40 batches):   811.5 KB ->  482.0 KB (-41%)
~8.2 KB saved per batch; scales with batch count.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* adjust changelog/releasenotes

* QueryLineWriter: guard null target, tolerate null query

Address PR review: Write threw NullReferenceException on a null query (via
query.Length). Restore the prior StreamWriter.WriteLine(null) tolerance by
treating null as empty (writes only the newline), and throw a clear
ArgumentNullException for a null target instead of an NRE.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* Rename QueryLineWriter to PooledStreamWriter, Write to WriteLine

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* Avoid per-row byte[] allocation in fixed-size binary reads (ClickHouse#669)

The read/deserialization path allocated a fresh byte[] per value via
ExtendedBinaryReader.ReadBytes(int). For fixed-size types the bytes are
consumed locally (BigInteger/IPAddress copy them out), so the array is
pure garbage.

Add an internal guaranteed-read ReadBytes(Span<byte>) helper and read
into a stack buffer instead:
- Int128/256, UInt128/256 (AbstractBigIntegerType): stackalloc + single
  new BigInteger(span, isUnsigned: !Signed) call, collapsing the old
  byte-by-byte unsigned path.
- Decimal128/256 mantissa (DecimalType): stackalloc + new BigInteger(span).
- IPv4/IPv6: stackalloc + new IPAddress(span).
- FixedString (default string mode): stackalloc for Length <= 256, else a
  pooled buffer, feeding Encoding.UTF8.GetString(span). The ReadAsByteArray
  mode is unchanged since that buffer escapes as the column value.

Behavior is unchanged (two's-complement decoding, IPv4 byte reversal).
In a 200k-row single-column read benchmark, managed allocations per read
dropped ~33-67% (e.g. UInt256 18.3 MB -> 6.1 MB) with fewer gen0 GCs; the
residual per-row alloc is the boxed result object (ClickHouse#509 territory).

Adds ReadDecodeBenchmark and server-free tests covering the FixedString
stackalloc/pool boundary, ReadBytes(Span) fill/underflow, and signed/
unsigned BigInteger high-bit decode.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* perf(read): override PeekableStreamWrapper.Read(Span<byte>) for direct span reads

ReadBytes(Span<byte>) routed through the default Stream.Read(Span<byte>),
which rents an ArrayPool buffer and copies per call — allocation-free but
adding rent/copy overhead in the tight per-row read loop. Override it to
read directly into the span, preserving the peeked read-ahead byte
semantics (including peek-past-EOF, which now returns 0). Addresses PR
review feedback (ClickHouse#448).

* test: cover PeekableStreamWrapper.Read(Span<byte>) edge branches

Adds direct tests for the two defensive branches in the new span-read
override: an empty span read while a byte is buffered from Peek (must not
write into the zero-length buffer or consume the pending byte), and a span
read after peeking past end-of-stream (must return 0, not 0xFF). The
Read(Span<byte>) method is now fully covered.

* docs: add XML comment to PeekableStreamWrapper.Read(Span<byte>)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* chore: remove ReadDecodeBenchmark from the PR

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>