Fix five correctness bugs in the binding engine + add regression tests - #8

Guards the B4 fix (PR #8): BindingContext.PropertyType must expose each
property's own type, not its declaring interface. Verified fail-first — with
PropertyType = propertyInfo.DeclaringType the assertion reports the property
type as the settings interface (ICaptureSettings) instead of System.String.

A second test documents that the fix did not remove access to the declaring
type: it stays reachable via SettingsType and PropertyInfo.DeclaringType.

Guards the B4 fix (PR #8): BindingContext.PropertyType must expose each
property's own type, not its declaring interface. Verified fail-first — with
PropertyType = propertyInfo.DeclaringType the assertion reports the property
type as the settings interface (ICaptureSettings) instead of System.String.

A second test documents that the fix did not remove access to the declaring
type: it stays reachable via SettingsType and PropertyInfo.DeclaringType.

Docs only; no code changes.

- FIX-PLAN.md (new): the prioritized, per-item fix plan from the three-part
  code review (architecture, tests, performance), with a progress banner.
- SESSION-HANDOFF.md: replace the stale modernization-era handoff with a
  current one covering the review, the fixes shipped this cycle
  (PRs #8/#10/#11/#12), key decisions (Validations and EqualityCompererCreator
  held for upcoming feature work), ranked next steps, and gotchas. The push/PR
  recipe is kept in private notes rather than inlined.

* Perf quick wins: enumerator, suffix match, env binder, type cache (Q1–Q4)

Five FIX-PLAN quick wins; Q5 was already resolved by B4 (#8), so this is Q1–Q4.

- Q1 SettingsCollection.GetEnumerator: was rebuilding a whole Dictionary via
  ToDictionary on every enumeration; now yields over the backing dictionary.
- Q2 SettingsTypesExtractor: replace Name.ToLower().EndsWith(suffix.Trim().ToLower())
  with EndsWith(suffix, OrdinalIgnoreCase) and hoist the trimmed suffix out of the
  per-type predicate. Also fixes a culture-sensitivity smell (ToLower was
  CurrentCulture). Regression test locks case-insensitive matching.
- Q3 EnvironmentVariableBinder: fast-path context.Key when there's no prefix and no
  formatter (skips the per-property StringBuilder), and collapse the Contains+indexer
  double lookup to a single IDictionary indexer read. Regression test covers the
  prefix branch.
- Q4 SettingsClassGenerator: cache the generated impl by interface Type in a
  ConcurrentDictionary instead of re-querying the module by mangled type name on every
  call. Regression test asserts the same interface returns the cached Type.

Build clean on net8.0 + net10.0; 55/55 tests green on net10.0 (was 52).

* Refresh handoff and fix-plan (P2 + docs merged, cleanup, Q1–Q4)

Bring the running status current: P2 (#18) and the docs tutorials (#20) merged,
the #8–#20 workstream branches pruned, and Q1–Q4 quick wins in flight (Q5 was
already resolved by B4). Flip the merged checklist items and re-rank next
priorities to P3 → P4 → P5, then engine tests (T7 concurrency race still open)
and architecture.

* Fix M1: namespace-qualify the generated impl type name (Q4 review)

Q4's Type-keyed cache exposed a latent collision: GenerateType derived the impl
type name from the *simple* interface name (GetNormalizeInterfaceName = Type.Name
minus leading I), so two settings interfaces sharing a simple name across
namespaces (Foo.ISettings + Bar.ISettings) both mapped to the same module type
name. Under the old name-keyed lookup the second silently reused the first's
(wrong) type; under the Type-keyed cache the second DefineType now throws and
aborts the whole scan.

Fix derives the impl name from the namespace-qualified FullName (sanitized),
kept deliberately separate from GetNormalizeInterfaceName — that helper also
backs SettingsOptions.SectionNameFormatter (the config section name), which must
stay simple-name-based. + a regression test generating two same-simple-name
interfaces. 56/56 green on net10.0.

* Add micro-benchmarks isolating Q1/Q3/Q4

The macro ScanBenchmark can't resolve the quick wins (they're <1% of the
IL-emit/populate cost). These isolate each changed hot path so the wins are
measurable and trackable:
- EnumerateBenchmark (Q1): enumerate a ~2000-entry ISettingsCollection.
- EnvBinderBenchmark (Q3): EnvironmentVariableBinder.BindPropertySettings fast path.
- GenerateTypeBenchmark (Q4): warm SettingsClassGenerator.GenerateType (cache hit).

Grants the benchmark assembly InternalsVisibleTo (Q4 uses the internal generator)
and references the Binders project (Q3 uses EnvironmentVariableBinder).