⚡ perf: memoize per-class plans; add depth guard and per-index each reporting

Profiling the validator showed roughly half of all validation time re-deriving
answers that cannot change — collectConstraints 22%, getOwnMetadata 12%,
getMetadataChain 9%, getProperties 4%, getMetadata 3%, plus 8% GC from the
allocation churn. The constraint predicates themselves were under 1%.

Decorator metadata is fixed once classes are declared, so the derived structures
are now memoized per prototype: the validation plan, the serialization plan, the
deserialization plan, and serializer/deserializer instances, which were being
constructed fresh for every property of every object. MetadataStorage carries a
version counter that invalidates the caches when metadata is written, so
registerDecorator after first use still works — covered by a test.

Measured against JSON.parse + JSON.stringify as a fixed reference:

  validate    (50 orders)   221.6 us -> 49.6 us   4.5x
  validate    (10 orders)    47.8 us -> 12.9 us   3.7x
  toInstance  (50 orders)   255.1 us -> 74.0 us   3.4x
  toInstance  (10 orders)    64.6 us -> 19.0 us   3.4x
  toPlain     (50 orders)   294.4 us -> 95.3 us   3.1x

Reliability, in the same pass:

- maxDepth option (default 64) on every mapping function, on validate(), and on
  configure(). All three engines recurse, so a payload nested thousands of levels
  deep could exhaust the call stack. Cycles were already handled; legitimate deep
  nesting was not bounded.
- each: true failures now name the element that failed ("failed at index 3"). A
  bad entry in a 200-item array previously produced a message that could not
  locate it. A message function now receives the failing element as args.value
  rather than the whole array; caller-supplied strings stay verbatim.

150 tests (up from 136), all green on the existing suite unchanged.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SAcqrz3FcadkYr3xG32CjK

CHANGELOG.md

@@ -5,6 +5,49 @@ All notable changes to this project are documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [Unreleased]
+
+### Performance
+
+Profiling the validator showed roughly **half of all validation time** was spent re-deriving
+answers that cannot change: `collectConstraints` (22%), `getOwnMetadata` (12%),
+`getMetadataChain` (9%), `getProperties` (4%) and `getMetadata` (3%), plus 8% garbage
+collection from the allocation churn. The constraint predicates themselves accounted for
+under 1%.
+
+Decorator metadata is fixed once classes are declared, so the derived structures are now
+memoized per prototype — the validation plan, the serialization plan, the deserialization
+plan, and serializer/deserializer instances (previously constructed fresh for every property
+of every object). `MetadataStorage` carries a version counter that invalidates every cache
+if metadata is registered late, so `registerDecorator` after first use still works.
+
+Measured on a customer record with nested address and orders, against `JSON.parse` +
+`JSON.stringify` as a fixed reference point:
+
+| Operation | Before | After | Speedup |
+| --- | --- | --- | --- |
+| `validate` (50 orders) | 221.6 us | 49.6 us | 4.5x |
+| `validate` (10 orders) | 47.8 us | 12.9 us | 3.7x |
+| `toInstance` (50 orders) | 255.1 us | 74.0 us | 3.4x |
+| `toInstance` (10 orders) | 64.6 us | 19.0 us | 3.4x |
+| `toPlain` (50 orders) | 294.4 us | 95.3 us | 3.1x |
+| `toInstance` (single) | 19.8 us | 8.4 us | 2.4x |
+
+### Added
+
+- `maxDepth` option (default 64) on every mapping function and on `configure()`. All three
+  engines recurse, so a hostile payload nested thousands of levels deep could exhaust the
+  call stack; it now raises a `JsonMappingError`. Cycles were already handled, but legitimate
+  deep nesting was not bounded.
+- `validate(obj, options?)` accepts options, so `maxDepth` applies to standalone validation.
+
+### Changed
+
+- `each: true` failures now report which element failed — `"... (failed at index 3)"`. A bad
+  entry in a 200-item array previously produced a message that could not locate it. A message
+  function now receives the failing element as `args.value` rather than the whole array;
+  caller-supplied string messages are still reported verbatim.
+
 ## [0.1.0] - 2026-08-03
 
 The first release with a working test suite. Everything below the "Fixed" heading was