fix: preserve histogram/exp_histogram temporality across msgpack round-trip

[truongnht]

Resolves #278

Summary

pack_header() in cmt_encode_msgpack.c only writes the aggregation_type meta field for CMT_COUNTER. Histograms and exp_histograms encoded to msgpack and decoded back always end up with CMT_AGGREGATION_TYPE_UNSPECIFIED, regardless of the temporality (delta/cumulative) they were created with — because unpack_basic_type_meta() in cmt_decode_msgpack.c never applies the decoded aggregation_type back onto those two types either.

This fix mirrors the existing CMT_COUNTER handling for CMT_HISTOGRAM and CMT_EXP_HISTOGRAM, on both the encode and decode side.

Real-world impact

This is the exact code path Fluent Bit's OTLP pipeline uses internally:

OTLP client → in_opentelemetry (decodes OTLP → cmetrics)
            → [internal msgpack buffer]
            → out_opentelemetry (decodes msgpack → re-encodes OTLP)
            → OTLP receiver (e.g. Prometheus)

Any OTLP histogram with delta or cumulative temporality passing through Fluent Bit loses that metadata in the msgpack round-trip. Downstream OTLP receivers that validate temporality (e.g. Prometheus with otlp-deltatocumulative enabled) reject the metric:

invalid temporality and type combination for metric "k6_http_req_duration"

Root cause, traced with evidence

• cmt_decode_opentelemetry.c / cmt_encode_opentelemetry.c (the direct OTLP codec) correctly round-trip aggregation_type for histograms — confirmed by reading the source.
• cmt_encode_msgpack.c / cmt_decode_msgpack.c (the internal msgpack codec Fluent Bit uses for its input/output plugin boundary) do not — confirmed by reading the source and reproducing locally.
• Reproduced against a live cluster: built a patched Fluent Bit v5.0.4 with this fix applied to the vendored lib/cmetrics copy, deployed it in place of the stock image, ran a k6 OTLP load test through the same Fluent Bit pipeline, and confirmed via direct Prometheus API query that k6_http_req_duration_milliseconds_bucket/_count/_sum (previously rejected) now land correctly with real data.

Test plan

Added tests/msgpack_temporality.c: round-trips a histogram, an exp_histogram, and a counter (as a control) through cmt_encode_msgpack_create → cmt_decode_msgpack_create and asserts aggregation_type survives.

• ctest passes: 18/18 (17 pre-existing + this new test), no regressions
• Verified the new test fails on histogram and exp_histogram (counter still passes, as expected) when reverting src/cmt_encode_msgpack.c / src/cmt_decode_msgpack.c only — confirms the test actually exercises the bug
• Verified the fix against a real Fluent Bit deployment with real OTLP traffic (see above), not just unit tests