docs: granularization / evidence-pack design

[ozgesolidkey]

Adds docs/GRANULARIZATION_DESIGN.md — the design for LOGAN's pre-LLM granularization layer.

What it specifies

• Principle: deterministic first, tokens last — LOGAN does the volume work natively; the LLM only sees the distilled artifact.
• Part 1 — Granularization layer: record stitching (line→record→event with stable unit_ids), persisted per-file structured index (persist what trend_fields computes), layered map-reduce summaries (whole-file → component → chunk → unit), addressability by unit_id + viewerLine.
• Part 2 — Evidence pack: one compact briefing (fields, level/component counts, crashes, top time-gaps, baseline deltas, anomaly flags) exposed as a single MCP call the agent fetches first.
• Part 3 — Cost-efficient tool use: work-by-reference + drill-on-demand, payload caps, proposer→deterministic-verify before pinning.
• What-exists-vs-new table + 5-step build order, and a tie-back to the AI-logging-framework idea (emit-time IDs vs analysis-time units).

Docs-only change.

🤖 Generated with Claude Code