Deterministic golden fixtures and a trace-format gate

[Jo5ta]

What

Makes golden trace-file generation byte-deterministic and adds a CI gate
that grows the fixture corpus only when the trace format actually changes.

Why

Golden fixtures let us test the decoders against known trace files, but until
now a fresh generation was non-deterministic (timestamps, pid, tid, span ids
vary per run), so nothing could tell "did the format change?" from "this is
just a new run". The result was a manual, easy-to-forget policy.

How

• frozen_info.c defines the three abstraction symbols the library uses to
  stamp entries (info_get_timestamp_ns/process_id/thread_id) with fixed
  constants. Linked ahead of the tracing static archive in the golden writer,
  it satisfies those references so the archive's info.o is never pulled —
  freezing timestamp, pid, tid, and (through the span-id salt derived from
  them) span ids. A generated trace is then byte-identical run to run. The
  core library is unchanged.
• One C++ writer (writer.cpp) exercises every tracepoint kind — printf,
  dump, dynamic (CLLTK_DYN_TRACEPOINT), spans, and fmt — into a single
  GOLDEN fixture. It's C++ because fmt is C++20-only; the other kinds compile
  identically from C and the on-disk format is language-independent, so one
  writer covers the whole decoder surface. Built by the golden-generate cmake
  target behind the CLLTK_GOLDEN option and the golden preset;
  generate.sh runs it in a per-arch Fedora container.
• The gate (scripts/ci-cd/step_golden.sh) regenerates the fixture at HEAD
  for both byte orders and byte-compares it against the newest committed
  golden, masking only the library-version field and its header crc
  (normalize.py, endianness-agnostic). Identical → the corpus already covers
  this format. Different → the format changed; regenerate_golden.sh writes a
  new version-named fixture (and refuses to overwrite an existing same-version
  baseline, so old-format coverage can't be silently dropped). Runs as its own
  CI job (qemu for le-aarch64 and be-s390x) and in run_all.sh
  (--skip-golden, auto-skipped when no container engine is present).
• ELF metadata fixtures reproducible from the same source. The previous
  committed golden-1.3.0-be-s390x.{o,so} blobs (for the clltk meta ELF test)
  had no source. They're now regenerated from writer.cpp compiled as a
  relocatable object + shared object (generate.sh --elf, golden-1.7.7-be-s390x.{o,so}).
  A separate cmake target keeps the format gate from building them, so an ELF
  build issue can't masquerade as a trace-format change.
• Deterministic 1.7.7 baselines for both byte orders. test_golden.py
  decodes every fixture with both decoders as the correctness oracle.

Testing

• Generation verified byte-deterministic across runs for both byte orders
  (le-aarch64 native, be-s390x under qemu).
• All fixtures decode to the expected messages with the Python and CLI
  decoders; the comprehensive fixture covers every tracepoint kind; clltk meta
  extracts the metadata from the committed s390x ELF objects cross-endian.
• Gate validated end-to-end: passes when the corpus is current; version-only
  changes are masked; a real payload change is caught.
• shellcheck (-x) clean; CMake presets / CI YAML validated.