v6.2: bulk generation, TimeProvider overloads, EF Core value generation

[buvinghausen]

Summary

SemVer minor (v6.2.0) — all changes additive. Implements the approved spec at
docs/superpowers/specs/2026-06-09-sequentialguid-v6.2-design.md.

• Bulk generation (Fill/FillSql/NewGuids/NewSqlGuids on GuidV7 and GuidV8Time, .NET 8+ assets):
  one timestamp capture, one counter-block reservation, one RNG call per batch — ~11× faster than a
  Guid.CreateVersion7 loop and zero-alloc. Loud ArgumentOutOfRangeException when a batch would
  exceed the counter space (2^26 v7 / 2^22 v8). Block reservation mirrors Interlocked.Increment
  post-increment semantics so bulk and single-call generation interleave without slot reuse
  (regression-tested with singles on both sides of a bulk call).
• TimeProvider overloads (.NET 8+) on every generation path, single-call and bulk.
• EF Core value generation: four public ValueGenerator classes plus
  UseSequentialGuidValueGeneration(bool sqlServerByteOrder = false) — a model-finalizing
  convention covering Guid, SequentialGuid, and SequentialSqlGuid primary keys (composite
  keys included). Explicit configuration always wins. Includes a HasSentinel(default) pin for
  struct keys: EF 8/9 compute sentinels via Activator.CreateInstance, which invokes the structs'
  generating parameterless ctors (fixed in EF 10 via GetUninitializedObject).
• MongoSequentialGuidGenerator gains a SequentialGuidType constructor; Instance stays v8
  (no silent behavior change). The spec's optional type parameter on RegisterMongoIdGenerator()
  was dropped — the instance already carries its type.
• Legacy GetInt32 bias fix: mask-and-reject sampling replaces GetNonZeroBytes + double-modulo.
• AOT smoke tests extended to bulk/TimeProvider/NodaTime; new EF Core AOT smoke project
  (public generators only — converters are internal and covered by build-time analyzers) wired
  into CI with per-exe exit-code guards. Native ILC publish verifies in CI.
• README: "Why not Guid.CreateVersion7?" section with measured benchmark table (honest about
  per-call latency; the win is monotonicity, SQL ordering, tooling, reach, and bulk throughput).

Minor deviations from spec snippets: TimeProvider tests live in their own file; bulk sort-order
tests use inline arrays; the bulk benchmark exercises Fill rather than NewGuids (fairer
buffer-reuse comparison); test field naming follows .editorconfig over spec pseudocode.

Test plan

• dotnet build — 0 warnings across all TFMs (warnings-as-errors)
• dotnet test — 31,605 tests, 0 failures, all TFMs (net11/net10/net9/net8/net472)
• Counter off-by-one regression test proven red against the bug, green after the fix
• Managed smoke runs PASS locally; native AOT publish + run for both smoke apps in CI
• BclComparisonBenchmarks run on Release; README table is verbatim measured output