feat(data): narrow an app's model to how that app uses it

[david-hudec-networg]

Stacked on #190 → #191 → #192 → #193. Only the last commit belongs to this one. GitHub cannot express a cross-fork stack, so the base has to be master. Review in order.

--detail full|minimal replaces the --columns-used / --scan-code pair from the closed #193, so neither flag ever ships. minimal answers how was this app built; full stays what the inputs declare.

Why the existing filter cannot get there

--columns-used resolves a reference by name against one global set. A single occurrence of createdon anywhere keeps the column on every table that declares it — 19 of 19 tables in one real app, from the views of three. That is a fair approximation for an author's column, which one table declares. It is useless for a platform column, whose name is identical on every table in the org.

Measured before this change: 252 of 830 emitted columns on one app were audit plumbing, 134 of 367 on another. --columns-used --scan-code had removed 65 by comparison.

Columns

A reference is now attributed to the table whose artefact made it — forms, views and charts by the folder they sit under, workflows by the PrimaryEntity in the manifest beside them.

Views and charts also get a FetchXML walk. A view selects columns of tables it joins to and names them in layoutxml through an alias that is sometimes the table name and sometimes opaque (alias="aa"), so crediting them to the folder's table would drop them from the table that owns them.

An artefact belonging to no single table — an app module, a sitemap, a plug-in source — can still only be matched by name. It therefore keeps an author's column and cannot rescue a platform one. Without that gate the original defect returns through the code scan: measured, 226 of 252 audit columns survived.

Where no artefact can be attributed to a table at all, its columns are classified instead. IsCustomField alone is wrong for this — 297 author-prefixed attributes across two real repositories carry 0, including primary name fields — so a publisher prefix read from the module's own Solution.xml overrides it. With neither signal present the column is kept rather than narrowed on no evidence.

Dropped whatever refers to them: logical columns, business-process-flow bookkeeping, and the base-currency twin — that last needs a name pairing, because both halves are marked as an author's.

statecode, statuscode, every primary key and every column an edge depends on are never dropped.

Tables

An N:N is admitted only when both its tables are in the app. On one app, systemuser alone contributed nine intersects and four far-side stubs.

The gate is read before the branch that builds the intersect, so the edge and both stubs are withheld together rather than leaving a dangling end — which is why the branch's unguarded dereferences need no restructuring.

A stub for a table an input does declare gets a new TableType.NotInApp rather than NotInSolution. The colour is chosen from the type, so that is where the distinction belongs � no extra field on Table, and DetailLevel never reaches a translator.

It applies wherever --app is given, not only under minimal: 13 of 14 stubs in one real app were declared as full entities in the same inputs, so the red "not in the solution" was untrue for almost all of them at either detail level. One AppScopeTests assertion from #192 changes with it.

Reporting

Every dropped column in one warning is unreadable at this volume. DroppedColumn carries a reason (NoReferenceFound, NoOwnArtefacts, PlatformPlumbing) and the command writes counts per reason through WriteData. That also clears the TXC028 this command was raising by returning a status envelope from a [CliReadOnly] command.

Four things this deliberately does not do

Each was built, then measured against both apps with a kill switch and removed:

	columns it bought
walk a view's FetchXML to credit a linked table's columns	0 � the token scan already credits the file to its own table, and those names appear in that table's own artefacts too
read a workflow's PrimaryEntity to credit its bundle	1, and it was owningbusinessunit � so not doing it is the better output
read IsCustomField	0 � the publisher prefix already answers it
classify an artefact-less table's columns rather than narrow them	0 � such tables are reported by name instead

What survived the same test: the unattributed-rescue gate (30 columns) and the .cs/.ts scan (14).

Cleanup carried here

Three duplications in code the earlier commits of this stack introduce, folded out rather than left for a follow-up:

• ColourFor hand-rolled FNV-1a in eight lines where SHA256.HashData is one. The palette changes; no diagram has been generated from it yet.
• The self-referencing check ran twice in the N:N branch, once on the synthesised column names and once on the tables' logical names.
• isSelfReferencing was read on the line after it was assigned.

Measured

Project and product repositories at develop:

app	detail	tables	columns	refs	enums	audit columns
ntg_projectmanagement	full	70	903	124	91	252
ntg_projectmanagement	minimal	46	538	94	82	42
ntg_administration	full	36	388	59	29	134
ntg_administration	minimal	22	218	39	28	23

Dropped: 311 columns on the first (248 unreferenced, 63 plumbing), 136 on the second (107, 29).

Five targets × two apps × two levels: exit 0, @dbml/core PARSE_OK, byte-identical on a second run. One table in 38 has no artefacts of its own (Letter) and is reported by name.

Tests

Two new files, 14 tests, no new dependencies. The regression that proves the change: a column only one table's form refers to is kept there and dropped on the other. Plus link-entity attribution, opaque-alias resolution through layoutxml, workflow PrimaryEntity, both fallback paths, the unattributed-cannot-rescue-a-platform-column rule, each categorical rule overriding a genuine reference, full leaving every design-only rule inert, and all five targets rendering after tables and columns are dropped.

AnEntityDeclarationDoesNotCountAsAReferenceToItsOwnColumns needed a form added: without one the table has no artefacts and takes the classification path, so the test no longer exercised what it claimed.

Known limits

• The both-sides N:N rule can drop an app's own bespoke association, not only generic plumbing — on one app, collab-workspace ↔ document-type. Nothing cheap distinguishes "shared platform table" from "this app's edge to a shared concept". --detail full shows them.
• sql / plainsql / edmx under minimal look complete but are not, and the formats cannot say so in-band. The help text does.
• createdon survives on all 19 tables of one app — correctly: each has its own product-layer views that sort by it. The rest of the audit block goes to zero.

🤖 Generated with Claude Code