feat(push): add time_series_classification category (backend#1054 WS2)

[LukasWodka]

What / Why

CLI workstream (WS2) of the time-series-classification epic (tracebloc/backend#1054): make data ingest --task time_series_classification work end to end against the sequence-grouped contract WS1 implements in the ingestor.

time_series_classification is the [redacted]-sepsis shape: a single CSV with fixed sequence_id / timestamp columns (Decision-2), where each sequence_id groups the timestep rows of ONE sequence and the label is constant within it — one class per whole sequence, and the platform counts the dataset in sequences, not rows (Decision-3).

Scope

• internal/push/category.go — one registry line: {ID: "time_series_classification", Family: FamilyTabular, Label: "Time-series classification", CLISupported: true, IsClassification: true}. Not RegressionClass → plain string label form (no label.policy). --task help, the interactive picker, and the accept-gate all derive from the registry, so no other listing edits were needed.
• Re-vendored internal/schema/ingest.v1.json + layout.v1.json via the scripts/sync-schema.sh mechanism: scripts/.data-ingestors-ref bumped to the WS1 branch head (c38c8ad). The registry↔schema parity tests (TestRegistryCoversSchemaCategories / TestRegistryWithinSchema) pin the new enum both ways and are green.
• internal/push/layout_contract.go — parse the contract's new grouping trait (Decision-4) and expose GroupingFor; all grouped preflight behaviour gates on the trait, never on a category id.
• internal/push/preflight.go —
  • label-diversity gate extended from the hardcoded tabular_classification to IsClassification(category) (mirrors the ingestor's is_classification wiring; covers TSC).
  • CheckSequenceSchemaColumns: sequence_id + timestamp must be declared in the schema (previews the ingest.v1 sequence-grouped conditional).
  • CheckSequenceRows: every timestep row must carry a non-empty sequence id (previews SequenceGroupValidator's null-id rule; NA sentinels count as null, pandas parity). Also yields the distinct-sequence count, echoed as an advisory note ("N sequence(s) … counts this dataset in sequences, not rows").
• Parity harness — 6 new tsc-* cases; goldens regenerated from the REAL WS1 validators (DATA_INGESTORS_DIR=/tmp/tsc_ws1, branch of data-ingestors#359). All pre-existing golden verdicts/values byte-identical. Two deliberate divergences documented in the manifest: tsc-label-flip (mid-sequence label flip — LabelConstantWithinGroupValidator rejects in-cluster, no local whole-group preview yet) and tsc-unsorted-timestamp (PerGroupTimeOrderedValidator, same gap). Candidate follow-up if burned uploads show up.
• README — category listing updated (also folds in the 5 text tasks the line had fallen behind on: 15 of 16 supported).
• Deliberately untouched (fixed-names Decision-2): spec.go, flags, interactive prompts — buildTabular already emits the correct string-label tabular spec for TSC.

Coupling

Requires tracebloc/data-ingestors PR tracebloc/data-ingestors#359 to land in the same release window (vendored schema comes from that branch — T16).scripts/.data-ingestors-ref pins that PR's head SHA; once #359 merges, bump the ref to the merge commit and re-run scripts/sync-schema.sh (the PR-branch SHA stays fetchable after a squash merge, so CI stays green either way).

Test plan

Verified locally:

• go build ./..., go vet ./..., gofmt -l clean, go test -race -cover ./... — all packages pass; scripts/coverage-floor.sh ok (push 84.6%).
• scripts/sync-schema.sh --check green against the new pinned ref (the exact CI invocation).
• Registry/parity/layout-contract pins: TestRegistryKnownCategories, TestSupportedCategories (now 15), TestRegistryCoversSchemaCategories, TestRegistryWithinSchema, TestRegistryMirrorsLayoutContract, new TestGroupingForMirrorsContract.
• New unit tests: TestCheckSequenceSchemaColumns, TestCheckSequenceRows, TestPreflightDataset_SequenceGrouped (dispatch-level: accept + sequence-count note, schema-missing reject, null-id reject, single-class reject, and time_series_forecasting unaffected), TestBuild_Tabular_PassesSchema/time_series_classification, TestBuild_TSC_SchemaConditionalRequiresSequenceColumns (proves the vendored conditional actually enforces Decision-2).
• TestValidatorParity: goldens regenerated from the real WS1 validators; all 33 cases pass; zero drift in pre-existing verdicts/values.
• Done-contract check: drove the real pipeline (DiscoverTabular → InferSchema → PreflightDataset → SpecArgs.Build → embedded-schema validation) over WS1's sample CSV (templates/time_series_classification/time_series_classification_sample_in_csv_format.csv, 6 patients / 30 rows / nullable lactate). Preflight accepts with the "6 sequence(s)" note; the produced ingest.yaml passes BOTH the WS1 branch's canonical ingest.v1.json (python jsonschema) and WS1's full real validator chain (map_validators("time_series_classification", …)) — schema + sequence-group + label-constancy + per-group order + numeric-columns + label-diversity all green.

Deferred to CI / dev:

• The kind-cluster integration e2e suite (nightly / e2e label) — not run locally.
• An actual in-cluster data ingest push against a live workspace — needs a dev cluster with the WS1 ingestor image; belongs to the epic's FR-on-dev stage.

Part of tracebloc/backend#1057 (epic tracebloc/backend#1054).

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches ingestion validation and couples release to data-ingestors#359; two parity divergences mean some bad TSC CSVs can pass local preflight but fail after upload.

Overview
Adds time_series_classification to data ingest so sequence-grouped tabular CSVs (fixed sequence_id / timestamp, one label per sequence) can be pushed end to end. The category registry marks it CLI-supported and classification-shaped; README now reflects 15 of 16 supported tasks.

Vendored contracts (ingest.v1.json, layout.v1.json) and scripts/.data-ingestors-ref are re-synced from the upstream WS1 branch: new enum value, schema conditional requiring sequence_id + timestamp, and a grouping trait on layout v2. GroupingFor reads that trait so grouped behavior is not hardcoded to a category id.

Preflight gains CheckSequenceSchemaColumns and CheckSequenceRows (null/NA sequence ids, sequence count note). Tabular label-diversity now gates on IsClassification (covers TSC, not only tabular_classification). Six tsc-* parity cases and goldens document two known gaps: mid-sequence label flips and per-group timestamp order are still accepted locally but rejected in-cluster.

^{Reviewed by Cursor Bugbot for commit 8c24142. Bugbot is set up for automated code reviews on this repo. Configure here.}