Perf quick wins Q1–Q4 (+ handoff/fix-plan refresh)

[guy-lud]

Implements the FIX-PLAN perf quick wins. Five items were planned; Q5 was already resolved by B4 (#8) — the dead string.Equals(x, null, …) checks in BindingContext are gone — so this is Q1–Q4. Each is a small, behavior-preserving optimization with a regression test where behavior could subtly shift.

Changes

• Q1 — SettingsCollection.GetEnumerator rebuilt a whole Dictionary via ToDictionary on every enumeration. Now yield returns over the backing dictionary — same pairs, same order, no per-enumeration allocation.
• Q2 — SettingsTypesExtractor matched the suffix with Name.ToLower().EndsWith(suffix.Trim().ToLower()) (two ToLower allocations per scanned type, under CurrentCulture). Now EndsWith(suffix, StringComparison.OrdinalIgnoreCase) with the trimmed suffix hoisted once out of the per-type predicate. Also removes the culture-sensitivity smell. (Regression test: a type ending …SETTINGS still matches the default Settings suffix — a case-sensitive EndsWith would miss it.)
• Q3 — EnvironmentVariableBinder allocated a StringBuilder per property and did a double lookup (Contains + indexer). Now fast-paths context.Key when there's no prefix and no formatter (the common case, zero allocation), and does a single IDictionary indexer read (returns null when absent; env values are never null when present). (Regression test added for the prefix branch.)
• Q4 — SettingsClassGenerator re-queried the module by mangled type name (Assembly.GetType(name.Replace("+","\\+"))) on every GenerateType. Now caches the generated impl by interface Type in a ConcurrentDictionary<Type, Type>. (Regression test: the same interface returns the cached Type.)
  • Note: this makes the cache thread-safe but does not close the pre-existing unsynchronized check-then-DefineType race — that stays a T7 concurrency item.

Verification

• Build clean on net8.0 + net10.0, 0 warnings.
• 55/55 tests green on net10.0 (was 52; +3 new regression tests).

Also

• Refreshes SESSION-HANDOFF.md + FIX-PLAN.md: marks P2 (Memoize property extraction, drop the O(n^2) dedup (P2) #18) / docs (Refresh the docs/ tutorials for the SimpleSettings rename and current API #20) / Q1–Q4 done, records the Fix five correctness bugs in the binding engine + add regression tests #8–Refresh the docs/ tutorials for the SimpleSettings rename and current API #20 branch pruning, and re-ranks next priorities to P3 → P4 → P5, then engine tests and architecture.

[guy-lud]

Update: M1 fix + micro-benchmarks (proof)

M1 (from review) fixed (656905e): the generated impl type name is now namespace-qualified in the generator only — GetNormalizeInterfaceName is left alone because it also backs the default config section name (SettingsOptions.SectionNameFormatter). Two settings interfaces sharing a simple name across namespaces (Foo.ISettings + Bar.ISettings) now generate distinct impls instead of aborting the scan. Regression test added; 56/56 green.

Micro-benchmarks added (f65890c) isolating each hot path — the macro ScanBenchmark can't resolve these (they're <1% of the IL-emit/populate cost). Before = master's impl, after = this branch (measured by overlaying the four production files, same benchmark code, ShortRun, [MemoryDiagnoser]):

Benchmark	Before (master)	After	Delta
Q1 Enumerate — 2000-entry ISettingsCollection	23.74 µs · 64.03 KB	8.66 µs · 88 B	2.7× faster, ~745× fewer bytes
Q3 BindFastPath — env binder fast path	29.16 ns · 152 B	10.99 ns · 0 B	2.65× faster, 0 alloc
Q4 GenerateWarm — warm GenerateType (cache hit)	74.81 ns · 224 B	2.33 ns · 0 B	32× faster, 0 alloc

Honest framing: these are per-operation wins on repeated paths (warm resolves, enumeration, per-property env binding). Cold startup is dominated by Reflection.Emit, so the macro scan number is unchanged — Q1–Q4 are allocation-hygiene + repeated-path speedups, not a first-scan speedup.

(Q2 is a culture-correctness fix — OrdinalIgnoreCase suffix match — with negligible, hard-to-isolate perf; no dedicated micro-benchmark.)