[holonsoftFounder]

Phase 2 complete. Three commits, each measured on its own.

1. Range violation coverage first (prerequisite)

Each of the five ConverterRangeViolationBehaviour values had exactly one test on one int field. Since the range check was about to move onto per-type accessors, a mistake affecting ushort but not int would have gone unnoticed — and these are read-direction paths, so the golden vectors are structurally unable to cover them.

AllNumericRangesPoco now carries all ten range-capable types (sbyte → decimal) in little- and big-endian variants, every behaviour asserted per field by name. Also pins that None behaves like IgnoreAndContinue rather than zeroing a field, and that an in-range value never invokes the handler. 72 → 83 tests.

2. Cached TypeCode

Type.GetTypeCode(...) ran per field per message in both directions. ConverterFieldInfo already had a private _fieldTypeCode but only populated it for range-checked fields — half-used, and the hot path recomputed it anyway.

Read −5.7% to −7.0%, Write LE −12.5%, Write BE −3.6%, strings −4% to −5%. Every row improves, which is expected here: both SetFieldValue and WriteFieldValueToArray called it, so no row should have stayed flat.

3. Typed accessors — no boxing in either direction

Action<T, object> on read and Func<T, object> on write boxed every value type once per field per message, and IsInRange(object) boxed again.

FastInvoke gained BuildTypedFieldSetter/BuildTypedFieldGetter. The pre-existing typed helpers only worked on PropertyInfo via Delegate.CreateDelegate, which cannot work for fields — a field has no get/set method. The new ones compile an expression that touches the field directly.

Method	boxing	typed	time	alloc
Read LE	118.86 ns	94.78 ns	−20.3%	344 → 104 B
Read BE	148.85 ns	115.31 ns	−22.5%	568 → 328 B
Read reused	114.23 ns	90.71 ns	−20.6%	288 → 48 B
Write LE	209.75 ns	204.25 ns	−2.6%	888 → 648 B
Write BE	333.25 ns	336.31 ns	+0.9%	1528 → 1288 B
Read w/ string	71.39 ns	73.42 ns	+2.8%	288 → 264 B
Write w/ string	182.10 ns	155.78 ns	−14.5%	472 → 448 B

Reading into a reused instance allocates 48 bytes instead of 288.

The write path barely moves in time because it is no longer dominated by the box. What remains is BitConverter.GetBytes allocating a byte[] per field, LINQ Reverse for big-endian, and the growing List<byte> — exactly what Phase 3 removes.

Enum.ToObject is gone too: enums are accessed as int and the compiled expression converts.

Cumulative against shipped 3.6.1 (bcea658)

	before	now	time	alloc
Read LE	118.09 ns	94.78 ns	−19.7%	344 → 104 B
Read reused	115.58 ns	90.71 ns	−21.5%	288 → 48 B
Write LE	245.33 ns	204.25 ns	−16.7%	888 → 648 B

Golden vectors unmoved throughout. 83/83.

🤖 Generated with Claude Code