Tracking: 0.4.1 Fabric reverse-engineering walk — capture/grade/YAML issues #44

Tracking

Filed from a DEKC 0.4.1 reverse-engineering walk of an Azure Fabric estate into a mixed PKC + SAC + DEKC second brain (2026-08-24). GitHub Latest was 0.4.1; Grok host install was still 0.3.2 (scripts used from the marketplace 0.4.1 cache).

Adversarial judges on that walk: coverage 0.76 (inventory OK), lineage 0.70 fail (typed edges), stream-job 0.71 (no fake streams), business 0.49 fail (no auto-promote, by design). dekc_grade.py crashed before automated RE scores.

Not duplicates of closed #3 (backslash doubling) or #7 (dekc_link CLI names). Those are related families.

Parser / grade (do these first)

• _parse_simple_yaml treats description: > as a nested map and swallows the rest of the frontmatter #26 _parse_simple_yaml treats description: > as a nested map (destroys PKC sources on _patch)
• dekc_grade.py TypeError when description is a dict (mixed bundles / bad YAML parse) #27 dekc_grade.py TypeError when description is a dict

Capture correctness (lineage-skeptic / coverage-skeptic)

• capture_table overwrites Schema contains so only the last table remains #28 capture_table Schema contains last-write-wins
• slugify collapses distinct Fabric names onto one file (pipe vs spaces, cross-workspace) #29 slugify collides pipe vs spaces / cross-workspace names
• Capture CLIs always stamp verified: true with no --verified flag #30 capture always verified: true
• capture_table --source invents sourced_from with no SQL or job evidence #31 --source invents sourced_from without SQL
• capture_lineage_path forces every hop through /tables/ and points feeds the wrong way #32 LineagePath tables-only + reverse feeds
• capture_ingestion_job always writes_to a Layer, so semantic refresh is modeled as a gold writer #33 IngestionJob always writes_to a Layer (semantic refresh as gold writer)
• capture_data_lake injects an unverified Bronze→Silver→Gold mermaid #34 DataLake unverified Bronze→Silver→Gold mermaid
• dekc_capture.py workflow still writes AGER type: Workflow after the 0.4.0 noun split #35 workflow subcommand still writes AGER Workflow after 0.4.0 noun split
• capture_semantic_model appends a vacuous body sentence that fails business-fidelity #36 vacuous SemanticModel body sentence
• Gold warehouse views captured via table land as type: Table; view CLI is easy to miss #42 gold vw* captured as type: Table

Graph / log / bulk

• write_knowledge emits a WriteEvent file per object with no consumer gitignore #37 WriteEvent file per object, not gitignored
• dekc_lineage.build_graph only reads links[], ignoring PKC rel: maps and markdown lands_as #40 build_graph ignores PKC rel: maps
• append_log writes one log.md line per captured object and drowns the human changelog #43 append_log floods log.md

Fabric / mixed-bundle (enhancements)

• dekc_walk.py only walks a SQL/parquet directory; Fabric REST / TDS profile is documentation-only #38 dekc_walk.py filesystem-only; Fabric REST profile is docs-only
• dekc_grade.py / dekc_doctor.py cannot grade a walk inside a mixed SAC+PKC+DEKC brain #39 grade/doctor not walk-scoped on mixed 15k-node brains
• Capture CLI has no fabric_item_id / workspace / datasetId / fabric_type fields #41 no fabric_item_id / datasetId / fabric_type on capture CLI

Suggested order

1. _parse_simple_yaml treats description: > as a nested map and swallows the rest of the frontmatter #26 then dekc_grade.py TypeError when description is a dict (mixed bundles / bad YAML parse) #27 (grade cannot run until parser or grade is defensive).
2. capture_table --source invents sourced_from with no SQL or job evidence #31 capture_lineage_path forces every hop through /tables/ and points feeds the wrong way #32 capture_ingestion_job always writes_to a Layer, so semantic refresh is modeled as a gold writer #33 (typed lineage pollution).
3. capture_table overwrites Schema contains so only the last table remains #28 slugify collapses distinct Fabric names onto one file (pipe vs spaces, cross-workspace) #29 Capture CLIs always stamp verified: true with no --verified flag #30 Gold warehouse views captured via table land as type: Table; view CLI is easy to miss #42 (identity and schema).
4. dekc_capture.py workflow still writes AGER type: Workflow after the 0.4.0 noun split #35 leftover noun split.
5. dekc_walk.py only walks a SQL/parquet directory; Fabric REST / TDS profile is documentation-only #38 Fabric walker so the next RE walk is not hand-REST + capture.