Claude-native telemetry: transcript evidence & compact export

Problem

Aggregates show where cost concentrated, not why. The "no AI slop" rule requires any finding be traceable to the actual transcript, and qualitative review benefits from a compact, readable rendering of prior sessions.

Capability

Reconstruct human-readable transcripts from the normalized stream and export a compact form for qualitative review and model-driven meta-analysis.

Behaviour

• Reconstruct conversation order from the parentUuid/uuid DAG; repair broken parent chains.
• Title sessions by the model-generated ai-title record, falling back to first prompt.
• Given a finding's sessionId + uuid, render the cited turn ± context.
• Compact Markdown export (configurable detail) that merges low-signal turns; small enough to feed back into a model.
• Date / session / project filters aligned to the sprint window.

Acceptance criteria

• Fixture reconstructs in correct causal order, including across a broken parent link.
• Title uses ai-title when present, else first prompt.
• Finding reference renders the cited turn ± context.
• Compact export materially smaller than full transcript and still coherent.

Privacy: exports stay under .logs/ (gitignored), never committed. Qualitative support for all dimensions. Tier 4. Depends on #36. PRD: docs/prd/features/08-transcript-evidence-export.md.

Start with docs/prd/features/08-transcript-evidence-export.md and inspect the normalized stream and its existing fixture or entry point. Verify the fixture reconstructs causal order across the broken parent link, selects the correct title, renders cited-turn context, and produces a coherent compact export. Confirm exports remain under .logs/.