Cache ObjectPType member dispatch

[chklauser]

Summary

• add bounded, DLR-style ObjectPType member call sites to interpreted functions
• emit the same call-site access from CIL-compiled functions, while preserving Prexonite's PType identity and recent-conversion-lock semantics
• retain four polymorphic rules, then switch permanently to a bounded megamorphic target that rebinds uncached misses
• add BenchmarkDotNet coverage for stable hits, expensive guards, fresh receivers, retained megamorphic hits, and deoptimizing five-shape cycles, including allocation and GC statistics
• make the CSharpier post-tool hook resolve JJ workspace roots correctly and update the README's reflection joke for the new reality

Why

ObjectPType previously repeated reflection discovery and Prexonite overload resolution for every member call. The call sites cache the result behind guards derived from the full PValue shape, rather than delegating binding to C#, so object overrides, type tags, and explicit conversions continue to follow Prexonite rules.

The four-rule cutoff bounds polymorphic-chain cost and retained memory. On the fifth distinct binding, a site keeps the four most recently bound shapes and becomes sticky-megamorphic; retained shapes still use cached bindings, while other shapes resolve directly without growing the DLR rule chain.

Benchmarks

BenchmarkDotNet 0.15.8, .NET 10, three launches, 15+ iterations, 2% maximum relative error, concurrent workstation GC:

CIL scenario	uncached	cached	cached allocation
zero arguments	346.95 ns	74.88 ns (4.63x faster)	344 B vs 1,232 B
explicit conversion lock	595.39 ns	189.99 ns (3.13x faster)	736 B vs 1,808 B
wide guarded call	856.13 ns	436.02 ns (1.96x faster)	1,808 B vs 3,152 B
megamorphic retained hit	427.04 ns	128.60 ns (3.32x faster)	448 B vs 1,376 B
megamorphic five-shape cycle	419.30 ns	218.32 ns (1.92x faster)	707 B vs 1,376 B
genuinely fresh receiver	453.59 ns	540.23 ns (1.19x slower)	1,952 B vs 1,640 B

The fresh-receiver case is the expected unfavorable case: it cannot reuse a binding and additionally checks the bounded retained set. Saturation reduces its earlier pre-cutoff regression from roughly 1.45x to 1.19x while preventing an unbounded guard chain.

Validation

• dotnet test -m:1: 1,446 passed, 3 skipped, 0 failed
• dotnet build -c Release Prexonite.Benchmarks/Prexonite.Benchmarks.csproj -m:1
• bash "$(jj root)/.codex/hooks/format-csharp.sh"
• full 16-case CIL BenchmarkDotNet suite with MemoryDiagnoser