v1.9.0

DeepDoc 1.9.0 replaces the LLM-discovers-structure planning model with topology-driven nav planning: call graph analysis pre-computes cohesive domain clusters before any LLM call, the LLM names and describes them, and flows are embedded inside their owning domain pages instead of a separate "Core Workflows" section.

Added

• Added deepdoc/planner/topology.py — build_topology_map() derives a TopologyMap from the call graph without any LLM involvement. It computes per-file indegree and BFS call-depth from entry points, groups files into TopologyCluster objects via BFS + Jaccard-based merging (threshold 0.40), and identifies a foundational cluster for shared infrastructure files (indegree ≥ 8% of repo size).
• Added topology_map: TopologyMap | None field to RepoScan; populated during Phase 2 scans immediately after call graph construction.
• Added _format_topology_clusters() to utils.py — formats topology clusters with entry files, key symbols, owned file counts, side effects, and external calls for the classify prompt.
• Added _build_named_clusters_str() to utils.py — merges LLM-assigned cluster names/sections/descriptions from the classify step with topology cluster file lists and call-graph signals into a rich context string for the propose step.
• Added _attach_flow_hints_to_cluster_buckets() in specializations.py — instead of creating a separate "Core Workflows" bucket, attaches flow_entrypoints, flow_id, flow_entry_kind, and sequence_diagram hints directly to the domain bucket that owns the flow's entry files.
• Added _build_section_depth_map() and _section_sort_key() in nav_shaping.py — order nav sections by topology cluster depth (entry-point-facing clusters appear first, foundational last) rather than by hardcoded section name lists.

Changed

• Classify step now sends pre-computed topology clusters to the LLM instead of a compressed file tree. The LLM names each cluster and assigns it a domain section, returning a cluster_names dict rather than per-file classification.
• Propose step now receives named_clusters (topology clusters enriched with LLM-assigned names/sections) instead of classification_summary + flow_candidates. Buckets are created from named clusters, not discovered from a compressed file-tree blob.
• _shape_plan_nav() now accepts an optional scan argument and uses topology depth to order domain sections; the old hardcoded _section_rank() is replaced by _section_sort_key() which only pins Start Here/Overview at the front and Testing/CI/CD/Supporting Material at the tail.
• _default_section_for_primary() no longer returns "Core Workflows" for backend_service repos — domain section names come from the LLM's cluster naming step instead.
• Duplicate _shape_plan_nav definition removed from heuristics.py; the authoritative version in nav_shaping.py is now used throughout.
• _ensure_flow_buckets and _expand_flow_bucket_ownership removed from the public planner API and replaced by _attach_flow_hints_to_cluster_buckets.

Fixed

• Fixed heuristics.py importing _section_rank from nav_shaping after the function was replaced by _section_sort_key.