[leandrodamascena]

Nested input conversion errors now identify the failing field, such as messages[0].content[1].text, while preserving exception classes and application causes. rboto.exceptions.get_conversion_context(error) exposes typed service, operation, path, expected rule, category and truncation metadata.

The shared core and generator build diagnostics while failures unwind. Map locations use entry positions, values stay out of automatic diagnostics, and context lengths are bounded. Builder-required fields come from explicit builder checks so fields with defaults are not reported as missing. All 34 services are regenerated. Python argument binding and deferred event-stream errors retain their existing handling.

Validation:

• 690 codegen/core tests and 131 compiled tests with S3, DynamoDB, Lambda and Bedrock Runtime; 55 service-dependent cases skipped.
• Strict mypy, Pyright, Ruff, workspace Rust tests, formatting, Clippy, documentation build and workflow lint.
• Five before/after processes per build: small input preparation 2.679 → 2.605 µs; 200 nested records 677.754 → 665.487 µs. This is an optimized adapter/core microbenchmark without networking, not an end-to-end release benchmark. Method and limits are in benchmarks/input_conversion.md.

Closes #70.