v0.9.5

Squeezes the row→Arrow→Parquet decode path and turns a three-engine A/B
benchmark (full vs chunked-parallel, on the same 2 M-wide and 10 M-narrow
tables, with perf and DBA-harm signals) into concrete, measured fixes. No
public-API breaks.

Performance

• perf(mysql) — SIMD-validate text columns on decode. The MySQL
  row→Arrow text-column append used String::from_utf8_lossy (scalar
  validation) while every other text path in the file — and the whole Postgres
  decoder — already used the simdutf8 helper. Switched the high-volume path
  to a simdutf8 fast path with a lossy fallback: 2.33× faster UTF-8
  validation on wide text (criterion mysql_utf8_text_append), byte-identical
  output. (Rides on this line's earlier opt-level = 3 and on-by-default write
  pipelining.)

Fixed

• fix(retry) — a statement-duration timeout is no longer retried. PG
  statement_timeout (57014), MySQL max_execution_time, and the MSSQL
  client-side cap were classed transient, so an unchunked full-scan that
  exceeds the budget re-failed identically up to 3× — measured at 20 m 22 s
  for 0 rows on a 2 M-row MSSQL full export. Now classified permanent (lock-
  wait / network timeouts stay transient) and the MSSQL message says to use
  mode: chunked or raise the budget. Same export now fails in 5 m 20 s
  (one attempt). Regression tests added.

Changed

• feat(init) — strategy defaults are cost-, engine-, and memory-aware.
  The A/B benchmark showed "parallel = faster" is false in general: a win on
  Postgres (2.1× + ~5× less VACUUM-horizon bloat), a regression on wide MySQL
  (already a fast sequential scan), and mandatory on SQL Server (a full
  statement times out). suggest_parallel now factors introspected
  avg_row_bytes and the engine (wide MySQL stays single-threaded), and emits
  the predicted peak RSS as a comment — a memory sweep established that peak
  RSS scales with batch_size × row width × worker count, not chunk_size.

Docs

• docs(bench) — REPORT_full_vs_parallel.md: the full three-engine A/B
  (perf + DBA-harm: longest open txn, PG xmin-horizon age, peak connections),
  the narrow-vs-wide reversal, the memory-footprint sweep, and the six
  corrections it justifies.