feat(dataset): port dry-run and impact from Go CLI

[beengud]

Closes #5

Ported surface

Adds two subcommands to the existing dataset route (existing list/view untouched):

• dataset dry-run <file.json> — Dry-runs saving a dataset pipeline; nothing is persisted. Prints the dataset that would be saved (Dataset: <name> (<id>)), datasets that would be rematerialized (Would rematerialize: <name> (<id>)), and compile/validation errors (Error in <name>: <text>). Exits 1 when any error datasets are returned.
• dataset impact <file.json> — Renders a table (NAME / ID / IMPACT) of downstream datasets affected by saving the pipeline. Compile/validation errors go to stderr. Supports --format json|csv / --json. Nothing is persisted.

Input JSON shape (matches the Go CLI):

{ "workspaceId": "...", "dataset": { "name": "..." },
  "query": { "stageQueries": [{ "stageID": "...", "pipeline": "..." }] } }

Schema-vs-Go corrections (the Go source did not match the published SDL)

The Go fork queried fields that do not exist in the published schema; these were rewritten against the real SDL (terraform-provider-observe):

Go source assumed	Published schema reality	What I did
saveDatasetDryRun mutation	No such mutation	Use saveDataset(..., dependencyHandling: { saveMode: PreflightDatasetAndDependencies }) — the schema's preflight/dry-run mechanism (matches operation/dataset.graphql)
errorDatasets { dataset { id name } errorText }	DatasetError { datasetId, datasetName, text } (no nested dataset, no errorText)	Select datasetName / text
getDatasetsAffectedByDatasetUpdate { affectedDatasets { dataset { id name } dependencyType } }	DatasetsAffectedByDatasetUpdateResult has only dematerializedDatasets, editForwardDematerializedDatasets, rematerializationCosts — no affectedDatasets/dependencyType	Build the impact table from dematerializedDatasets + editForwardDematerializedDatasets (dataset { id name }); the IMPACT column reflects which list a row came from. dependencyType is not exposed by the schema and could not be reproduced
dataset: { name } mapped directly	DatasetInput.label: String! (no name field)	Map input dataset.name → DatasetInput.label
query.stageQueries[].stageID	MultiStageQueryInput requires outputStage: String! + stages: [StageQueryInput!]! where the field is id (not stageID) and input: [InputDefinitionInput!]! is required	Map stageID → id, set input: [], derive outputStage from the terminal stage

The impact command issues a second preflight saveDataset to obtain errorDatasets, since getDatasetsAffectedByDatasetUpdate does not return them.

Verification

• Codegen (scoped to src/gql/dataset/**): ✔ Generate — both operations validate against the published SDL.
• Typecheck (scoped tsconfig): no errors in src/commands/dataset or src/gql/dataset source files (pre-existing unrelated errors from missing rest/generated/other resources' types are out of scope).
• Lint/format: eslint + prettier clean on all new/edited files.
• Tests: bun test src/commands/dataset/{dry-run,impact,input}.test.ts → 14 pass / 0 fail, 32 assertions across 3 files.

🤖 Generated with Claude Code