The Metrics SDK hardcoded a 2000 cardinality limit and silently discarded any user-supplied override from declarative YAML config.

Wired a configurable limit through MetricReader and CollectorHandle end-to-end, fixing a sentinel-value bug along the way; a follow-up fix (#4314) closed a gap where histogram views setting only a cardinality limit were silently rejected.

Get("")/Set("") silently touched the wrong slot instead of normalizing to "_", a spec violation - fixed alongside documenting the caching contract's string_view lifetime guarantees.

minimum_severity and trace_based were missing from LoggerConfig's YAML configuration path, so file-configured loggers silently ignored severity filtering.

Values were stored raw instead of percent-decoded per the W3C Baggage grammar the spec defers to, so key=hello%20world produced the literal string instead of "hello world".

Caught and fixed a real operator-precedence bug during review: << binds looser than +, so combining hex nibbles with + rather than | would have silently miscomputed the decoded byte.

Swapping deprecated C headers for their C++ equivalents left calls like memcpy, strcmp, and strtod unqualified, relying on implementations happening to expose them globally, which the standard does not guarantee.

Added std:: to those call sites across api, sdk, exporters, ext, examples, and functional, plus the includes that had relied on transitive inclusion, leaving vendored code untouched.

predicate_factory.h used uint8_t without including cstdint, compiling only because another header in the same translation unit happened to pull it in.

Added the missing include directly to the header that needs it, so it builds in isolation and under standard libraries without that transitive include.

Closed out a long-running clang-tidy tracking issue by wrapping the last four file-local symbol groups in anonymous namespaces, reducing ODR risk on symbols with no header exposure.

Swept ~90 files from legacy C headers to their C++ equivalents, fixing IWYU's mappings first so it would stop recommending the headers being removed; excluded vendored code a reviewer flagged as out of scope.

Filed a scoped follow-up (#4356) for the std:: qualification gap the header swap exposed, rather than scope-creeping it into the same PR.

When a view set aggregation_cardinality_limit without an explicit aggregation, AddView always built a plain AggregationConfig. For histogram and exponential-histogram instruments that type didn't match, so the whole view was dropped with a warning and the limit never applied.

AddView now resolves the instrument's effective aggregation type the same way the view registry does, so the configured limit takes effect.

WriteValue stored a log record's body/attributes as given, and Export() dumped them via nlohmann::json with its default strict error handler - any non-UTF-8 byte (a truncated multibyte sequence, a differently-encoded payload) threw type_error.316 inside a noexcept function, turning std::terminate() into the actual failure mode for one bad log record.

Switched to error_handler_t::replace, which substitutes U+FFFD for the invalid bytes so the record and the rest of the bulk batch still export instead of crashing the process.

A record created while the logger was disabled was a no-op record, but enabling the logger before emit let it reach two unconditional static_casts, which is undefined behavior. The issue's dynamic_cast fix doesn't build under the project's no-RTTI CI job.

CreateLogRecord now returns an empty MultiRecordable while disabled, which is a real Recordable, so both casts are always safe and the record is dropped without reaching a processor.