Substring optimizations: eliminate hot-path allocations (audit top 8)

[christophwille]

Implements the first eight recommendations of a repository-wide Substring usage audit (105 production call sites classified by call frequency and by what the substring result feeds into, plus pattern micro-benchmarks). Each commit corresponds to one recommendation, ranked by (frequency x allocations saved) / effort:

1. AssignVariableNames.SplitName - the digit tail is already validated by the preceding loop, so the Substring + int.TryParse pair is replaced with an inline, overflow-guarded digit accumulation. Runs per variable and per reserved-name registration of every decompiled method.
2. AbstractSearchStrategy - operator-prefix stripping and (for ~ fuzzy terms) lowercasing now happen once per search run in the constructor instead of per candidate name per keystroke; the noncontiguous matcher compares spans with per-char ToLowerInvariant. This was the largest allocation source in interactive search (the ~ branch allocated two full ToLower() strings per entity). New IsMatchTests pin the +/-/=/~ semantics.
3. ReflectionHelper / TopLevelTypeName - the arity after the backtick is parsed in place (netstandard2.0 has no span int.TryParse) and namespace/name are each cut exactly once, removing the throwaway intermediate per generic reflection name parsed. Micro-benchmark of this pattern: 74.2 -> 24.1 ns/op, 386 -> 0 B/op.
4. MethodBodyDisassembler.WriteOpCode - ldarg.N/ldloc.N/stloc.N shortcut forms render their digit and param_/loc_ reference keys from static tables instead of two allocations per instruction.
5. TextWriterTokenWriter.EscapeIdentifier - returns the original string instance when nothing needs escaping (the overwhelmingly common case; previously it always built a StringBuilder plus a fresh result string per identifier token) and appends surrogate pairs as two chars instead of a 2-char substring.
6. CleanUpVariableName chain - all prefix/suffix cuts on the per-variable naming path are now slices; ContainsNonPrintableIdentifierChar and IsValidName gained span overloads, and only the final lowered name is materialized (single allocation). IsKeyword deliberately keeps its string parameter: it is called on that final string anyway, and the keyword HashSet has no span lookup on netstandard2.0.
7. IL-with-C# view - ISmartTextOutput gained Write(ReadOnlySpan<char>) as a default interface method (net10.0), overridden in AvaloniaEditTextOutput to append straight into its buffer; the highlighted-comment writer slices each source line instead of cutting up to four substrings per emitted line. The overload is on ISmartTextOutput rather than ITextOutput because the decompiler library targets netstandard2.0 (no DIMs), where a new interface member would break external implementers.
8. MetadataMethod - the explicit-interface-implementation op_ check tests a span slice first and only allocates the short name for actual operators, instead of for every static non-generic method containing a dot.

Deliberate behavior changes (both called out in the respective commit messages):

• Fuzzy search matching (~) now lowercases with the invariant culture instead of the current culture.
• Reflection-name arity parsing accepts only plain ASCII digits; suffixes like Foo`+1 that int.TryParse tolerated are now rejected (not legal reflection names).

Verification: full ILSpy.sln test suite (including the complete PrettyTestRunner golden suite, the disassembler golden tests, and the headless-Avalonia UI tests); new unit tests for the search matcher, reflection-name/arity parsing, and identifier escaping.

🤖 Generated with Claude Code

[christophwille]

Original analysis document: SubString.html