0.4 — OTLP traces: Jaeger, Tempo, Quickwit, OpenObserve

[DeviousCardi]

Summary

Part G of PLAN-1.0.md: a new otlp-traces suite (15 checks) run against four
backends — Jaeger, Grafana Tempo, Quickwit, OpenObserve — each of which
represents a span in a genuinely different shape than OTLP's own.

All 15 cases pass on all 4 columns (60/60). 183 unit tests, 0 clippy warnings.

Runner changes (generalize past traces)

• Absolute-URL support in Runner::send() and docker::wait_ready(): a
  rendered request path starting with http:///https:// is used verbatim,
  bypassing base_url. Needed because Jaeger and Tempo split ingest and
  query across two ports of one container. Regression-tested with two
  separate stub servers so neither endpoint can fall back to the other.
• Container.extra_ports / Container.config: publish additional
  container ports, and mount an inline config file at a fixed path, for
  settings with no CLI-flag equivalent. Tempo needs this — its OTLP receiver
  binds to 127.0.0.1 inside the container unless the config sets an
  explicit 0.0.0.0 endpoint, confirmed only via docker logs, not the API.
• export_report() made protocol-aware: it was comparing
  request_encoding == "otlp-json" literally, silently wrong for every OTLP
  metrics/traces JSON export, and its protobuf decode path was hardcoded to
  the logs response type — a genuine traces/metrics partial-success report
  decoded to None and was invisibly dropped before this fix.
• field_of() fallback chain extended: attributes.<key> searches the
  record's own OTLP-shaped attributes array; any other dotted field with no
  adapter mapping is tried as a JSON pointer.

A regression found and fixed along the way

An earlier edit meant to add trace_id_base64 to vars_for() silently
failed to save (a multi-assert Python script whose last assert failed, so the
final file write never ran) while a separately-applied edit to read_back()
for the same feature did save — masking the omission. The always-empty
trace_id_base64, combined with first_record()'s .contains(run_key)
check ("".contains("") is true in Rust), spuriously matched the first
record in any response. Fixed by inserting the missing variable and adding a
permanent guard: first_record() now refuses to match an empty key, with a
regression test.

Adapter notes (confirmed against running containers)

• Tempo answers real OTLP JSON on read-back, except trace/span IDs come
  back base64 (protobuf JSON's bytes-field encoding) where this project's
  encoder uses hex — compared as bytes, not equal strings.
• Jaeger answers only its own query model: microsecond timestamps, and a
  span event becomes a logs[] entry.
• Quickwit and OpenObserve flatten spans into search columns;
  OpenObserve stores events as a JSON string, unreachable by any field
  pointer.

event-timestamp-outside-span is match: present rather than exact
because no store here keeps OTLP's own field shape for a span event — the
check records where each one put the data.

🤖 Generated with Claude Code

https://claude.ai/code/session_018k65nFSzuwsHwYnpSHNaeK