feat(jsonlogger): schema presets, structured errors, redaction, term caps

[Taure]

Summary

Tier 1 modernization of nova_jsonlogger to bring it in line with what shops expect from a 2026 structured-JSON logger.

• Schema presets via #{schema => nova | ecs | otel | gcp | datadog} config:

  • ECS — @timestamp, log.level, message, trace.id, span.id, error.type / error.message / error.stack_trace, log.origin.* for source location.
  • OTel — Timestamp (nanosecond), SeverityText + SeverityNumber (1-24 per spec), Body, TraceId, SpanId, exception.*.
  • GCP — severity (uppercase), message, logging.googleapis.com/trace, logging.googleapis.com/spanId, structured sourceLocation.
  • Datadog — status, message, dd.trace_id, dd.span_id.
  • nova (default) preserves the current shape and now emits RFC 3339 timestamps by default.

• Structured error object from crash_report (SASL convention) or class / reason / stacktrace meta — #{type, reason, message, stacktrace => [#{mfa, file, line}]} rendered to each schema's conventions.

• Body-path redaction — #{redact => [[req, headers, authorization], [user, password]]} walks nested maps and replaces matched paths with <<\"[REDACTED]\">>. Missing paths are no-ops.

• Term-size guards — max_term_size and max_string_length (default 8192 bytes) cap ~0p-formatted output and binary/string values, appending \"...[truncated]\" to truncated payloads. Prevents a 10 MB term from blowing the log line.

• Existing key_mapping, format_funs, meta_with, meta_without, new_line[_type] keep working and now run after the schema layer so users can override anything the schema rendered.

Trace correlation

trace_id / span_id are picked up from process metadata. Upstream opentelemetry_nova PR #10 (already merged) writes them in pre_request, so once it's pulled the keys land under each schema's conventional names automatically. No OTel dependency in nova_jsonlogger itself.

Pipeline

merge_meta -> extract_error -> redact -> apply_schema ->
apply_key_mapping -> apply_format_funs -> pre_encode (with caps)

Behaviour change

Meta time (microseconds) is now emitted as an RFC 3339 binary by default under every schema. Callers using format_funs => #{time => fun(Int) -> Int end} should switch to a binary-accepting fun, or pass #{time => fun(B) -> B end} to keep the value untouched.

Test plan

• 35 eunit cases (17 new) all passing
• Full nova eunit suite (364 tests) passes
• xref clean
• dialyzer clean
• New tests cover: each schema's rendering, source location (ECS + GCP), severity number mapping, stacktrace extraction, error rendering under ECS + OTel, top-level + nested + missing-path redaction, binary + term + string caps.

[burbas]

Think it looks good except for the hard-coded thoas-lib usage :-). Fix that and it's A-Ok!

[burbas]

Here we should call nova to see which json-lib to use.

[Taure]

Good catch. Removed the hard-coded json-lib usage: introduced a shared ?JSONLIB macro in include/nova.hrl ((nova:get_env(json_lib, thoas))) and routed nova_jsonlogger encoding plus all the tests through it, so the library is fetched from config in one place rather than resolved at each call site. Default lib is unchanged. eunit (364) + xref + dialyzer all green.

[burbas]

Very picky now, but I think you can call the module directly instead of adding a macro for it :)

[burbas]

Rewrite function and explain where trace_id and span_id comes from

[burbas]

Looking at this now and I'm a bit unsure why we return a lambda here. Could we not just match in the header and have to clauses? Seems unnecessary to create a lambda here imho.

[burbas]

I'm a bit unsure where trace_id (And perhaps span_id) is set? I can see that the code tries to do call rename/3, but it's just a rename op - not a set. I might have missed this, but where is this done? :)

[Taure]

Good questions, both addressed in aad5296.

trace_id/span_id: you're right that rename/3 only relocates, never sets. That's deliberate - the formatter doesn't generate trace context. The keys arrive in the log record's metadata, merge_meta surfaces them, and rename/3 moves them to each schema's conventional name (trace.id, TraceId, ...), no-op when absent. The producer lives upstream: opentelemetry_nova #10 writes the hex ids into logger metadata in pre_request, or you call logger:update_process_metadata/1 yourself. nova has no OTel dependency, so I reworded the moduledoc to stop implying it's automatic and added a comment at the schema layer. Two tests now lock it: ids surface from metadata, and nothing is invented when they're absent.

normalise_crash_report: dropped the lambda and the polymorphic report_value/4. error_info/1 pattern-matches the {Class, Reason, Stack} tuple directly in the head, with a small take_into/3 for registered_name/pid. Added crash_report tests too - that path had none.