package:bench_press v0.4.0

• Breaking (library exports): package:bench_press/bench_press.dart now
  exports only the benchmark-authoring (Benchmark, AsyncBenchmark,
  BenchmarkGroup, BenchmarkMatrix, BenchmarkVariant, BenchmarkConfig,
  Throughput, ByteThroughput, ElementThroughput, Blackhole), suite
  entrypoint (mainBenchmark*), and .report() result (BenchmarkResult,
  BenchmarkMetrics, CalibratedBatch, WarmupResult) APIs. Internal CLI,
  subprocess runner, compiler, statistical helper, and telemetry schema types
  that were previously re-exported in 0.3.1 are now internal to lib/src/.
• Breaking (config): removed the matrix.entrypoints key from
  bench_press.yaml. It never chose which file ran: each value re-ran the same
  discovered benchmark files under a made-up entrypoint coordinate, so the
  matrix table compared identical runs. Pass benchmark files as positional paths
  to bench_press run instead. A leftover entrypoints: key is ignored.
• Breaking (behavioral): ByteThroughput.formatRate now scales byte rates
  by decimal 1000 (KB/s, MB/s, GB/s) instead of 1024, matching
  hardware memory bandwidth, network I/O conventions, and ElementThroughput.
• Breaking (behavioral): bench_press diff and bench_press run --diff now
  pair cells by full key (name, target, and every coordinate, including group)
  instead of by name and target alone. A name reused across groups was
  previously compared against the first match in the other run. The Benchmark
  column now shows coordinates, for example encode (group=A). Cells present in
  only one run are listed in an Unmatched note, including when nothing pairs.
• Breaking (behavioral): an explicitly configured Dart SDK (sdk matrix
  axis in bench_press.yaml) is now authoritative and fails fast with a
  specific diagnostic explaining why the path was rejected instead of silently
  falling back to dart on PATH.
• BenchmarkGroup.report, BenchmarkMatrix.report, and mainBenchmarkSuite
  now interleave measurement trials across group variants in ABBA BAAB rounds
  after every variant has finished setup, warmup, warmupComplete, and batch
  calibration. Previously, all trials of variant A ran before variant B
  started, so a linear thermal ramp or a short host contention window during B
  produced a false speedup or regression while both variants still passed the
  within-variant isStable gate. When maxTrials is set, lockstep rounds
  continue across all variants in the group while any variant's CV exceeds 5%.
• bench_press run now honors defaults.output in bench_press.yaml, both for
  saving results and for the file run --diff looks up. An explicit --output
  or --save still takes precedence, and benchmark_results.json remains the
  fallback. A leading ~ expands to the home directory. The value must be a
  non-empty string; anything else is now a config error, where it was previously
  ignored. bench_press report and bench_press diff do not read it; pass them
  the path.
• Added --pin-cpu <cpu-list> to bench_press run to pin benchmark
  subprocesses (VM JIT, AOT, Node.js, and D8) via taskset -c on Linux.
• Added Fieller confidence interval and isStable speedup gating (gate: true
  by default, configurable via --[no-]gate in bench_press run, report, and
  diff). Unstable or unbounded-CI comparisons render as unresolved and are
  excluded from geometric mean rollups; a resolved comparison whose 95% CI
  contains 1.00x renders as ➖ ⚪ Neutral.
• Added Markdown report warning banners to flag benchmarks whose declared
  Throughput.bytes rate exceeds physical memory bandwidth (100 GB/s) or
  whose latency remains invariant (0.9x–1.25x) across >=8x payload size
  spreads for a shared benchmark name across groups.
• Derived post-warmup calibration batch sizes directly from steady-state warmup
  convergence latencies (recorded in warmup.estimated_op_ns), and added a
  Markdown report footnote listing any cell whose trial median moved more than
  25% from the warmup estimate that sized its batch.
• Updated bench_press run and bench_press validate to discover and execute
  all positional file and directory paths supplied on the command line rather
  than silently ignoring arguments after the first path.
• Invoked warmupComplete() lifecycle hook during BenchmarkVariant execution,
  forwarded throughput and warmupComplete() for Benchmark and
  AsyncBenchmark instances run via mainBenchmark*, and preserved
  programmatic BenchmarkConfig values unless explicitly overridden by CLI
  flags.
• Fixed Markdown suite reporting so suites mixing standalone benchmarks and
  BenchmarkGroup variants collate grouped variants into a single multi-row
  ### Group: ... comparison table with the baseline ordered first and a
  geometric mean summary footer, and stopped wrapping Markdown tables in
  mdformat off / mdformat on HTML-comment guards.