test: add representative coding-agent evidence fixture

[harshitethic]

Summary

Adds a versioned, synthetic representative session fixture as a reusable regression input for #250.

Scenario covered

The fixture contains:

• explicit session start/end
• a redacted user prompt
• file read and write events
• a shell test command that fails
• an error linked to the failed call
• a retry linked to the first attempt
• a successful recovery result
• an agent decision
• a human correction annotation
• one deliberately orphaned tool result representing a known capture gap
• a final assistant response

Artifacts

• tests/fixtures/representative_session/events.ndjson — canonical TraceEvent input
• tests/fixtures/representative_session/annotations.jsonl — human review overlay
• tests/fixtures/representative_session/expected.json — deterministic IDs/count/gap expectations
• tests/test_representative_fixture.py — schema + scenario regression coverage
• docs/representative-fixture.md — trust boundary and refresh procedure

Why the gap is intentional

The fixture includes an orphaned tool result on purpose. Review UI/evidence-health work must surface missing relationships instead of inventing a parent and presenting a falsely complete timeline.

Scope

This is a focused foundation for #250 rather than a claim to close every consumer listed there. Follow-up PRs can wire the same fixture into evidence health, share/export, OTLP mapping, and the local review UI without each subsystem inventing different sample data.

Validation

The test parses the fixture through the repository's own TraceEvent.from_json() and Annotation.from_json() models and asserts the required failure/retry/recovery, redaction, annotation, relationship, and intentional-gap invariants.

I could not execute the suite locally because the connected development machine is offline; upstream CI is the verification gate for this branch.

Refs #250.