feat(pipeline)!: transaction-aware Writer and the Executor→Reader rename

[ddeboer]

First slice of #534 (search as a Configurable Pipeline instance): the @lde/pipeline core changes everything else builds on. Design captured in ADR 6.

Rename Executor → Reader (commit 1)

The input unit is now named for the read side of the familiar Reader/Writer I/O pair (the Stack’s source/sink): Executor → Reader, execute() → read(), SparqlConstructExecutor → SparqlConstructReader, ExecutorContext → ReaderContext, AttachedExecutor → AttachedReader, StageOptions.executors → readers. Mechanical; @lde/pipeline-void re-exports the renamed types, @lde/pipeline-shacl-sampler adapts internally.

Transaction-aware Writer (commit 2)

Writer becomes a factory of per-run transactions, generic over its payload (Writer<Item = Quad>):

interface Writer<Item = Quad> {
  openRun(context: RunContext): Promise<RunWriter<Item>>;
}
interface RunWriter<Item = Quad> extends DatasetWriter<Item> {
  flush?(dataset: Dataset): Promise<void>;
  reset?(dataset: Dataset): Promise<void>;
  commit(): Promise<void>;
  abort(error: unknown): Promise<void>;
}

• Pipeline.run drives openRun → write* → commit/abort uniformly and never branches on the writer’s update mode: an alias swap is a private step of a Blue/green writer’s commit(), a deletion sweep of an In-place writer’s.
• RunContext carries runId, startedAt, selectedSources() (complete by commit time, including datasets skipped as unchanged – the input to a registry-membership sweep) and the provenance store.
• Stages write through the narrower DatasetWriter, so a stage can never commit or abort the run.
• Lifecycle-free destinations wrap a per-dataset write with perDatasetWriter() – deliberately no second TransactionalWriter interface.
• A stage failure stays a per-dataset outcome (run still commits); a run-level failure (e.g. selection dying mid-iteration) aborts every fanned-out branch, even when one abort fails.

Ridden-along fixes

• Per-run state (open files, cleared graphs) now lives in the run, so re-running a pipeline on the same writer instance replaces output instead of appending – previously a latent bug.
• Chained (sub-stage) scratch FileWriters are now flushed before their output path is resolved – previously the chain resolved a final path that only materialized on a flush nobody issued.
• FileWriter.commit finalizes files still open; abort discards temp output, never leaving a truncated final file.

Breaking changes

Both commits are marked !. Custom writers implement openRun (or wrap with perDatasetWriter); Executor call sites rename. The Dataset Knowledge Graph pipeline adapts on its next @lde/pipeline upgrade.

Part of #534 (slice 1 of 4; next: BlueGreenRebuild/InPlaceRebuild writers in @lde/search-typesense).