fix(preflight): TSC group-integrity parity — mirror di#359 label-constant + time-order validators (#218)

[LukasWodka]

What

Closes #218. The CLI's time_series_classification local preflight previewed only column-presence + null-sequence_id rows — so a mid-sequence label flip or per-group unsorted timestamps passed the local dry-run but were rejected in-cluster. This mirrors di#359's two TSC group validators as a contract-locked preview (Principle 6: the ingestor owns the logic; the CLI mirrors it under the parity harness).

What it adds

CheckTSCGroupIntegrity (internal/push/preflight.go) — one pass grouped by sequence_id, previewing:

1. label-constant-within-group (≙ LabelConstantWithinGroupValidator): one label per sequence. Null-id rows dropped (pandas dropna); a null label counts as its own value; numeric-label collapse only when every non-null cell is numeric (so "1"/"1.0" isn't a false flip); single-row / already-constant groups never flagged.
2. per-group time-ordered (≙ PerGroupTimeOrderedValidator): monotonic non-decreasing per sequence (ties OK, interleaving OK). The numeric branch is mirrored (NA/unparseable → reject); the TIMESTAMP/date branch is intentionally a documented safe under-preview (accept) — reproducing pandas' permissive, locale-ambiguous date parse in Go would risk over-rejecting a date the cluster accepts.

Strictly TSC-scoped: the sole caller gates on the vendored layout contract's grouping trait, so no non-grouped / non-TSC dataset is touched.

Parity + deployment safety

• .data-ingestors-ref already pins the di#359 merge (7b4ecac); goldens for tsc-label-flip / tsc-unsorted-timestamp / the new tsc-numeric-label-na-sentinel case are generated by driving the real Python validators (not hand-authored). sync-validator-goldens.sh --check is green.
• di#359 / TSC is not in the deployed prod ingestor (v0.7.0 held it back) — this is develop-ahead, and the checks fire only for TSC, so they can't out-strict any deployed non-TSC path.
• NA-sentinel fidelity: the two grouped validators read via plain pd.read_csv (pandas default na_values), so the checks use pandas' exact default STR_NA_VALUES (19 tokens), not the CLI's curated coercion set — otherwise a numeric label column pandas NA's out would be a false flip (over-reject). Whitespace handled faithfully (raw cells; trim only inside the numeric parse); numeric collapse uses float64 (documented int64>2⁵³ edge — under-rejects, the safe direction).

Tests

• TestValidatorParity green incl. label-flip / unsorted-timestamp (reject) + the NA-sentinel case (accept).
• 14-case TestCheckTSCGroupIntegrity (single-row, ties, interleaving, numeric collapse, null-label flip, invalid numeric time, null-id drop, date-branch benign skip, padded-object flip, NA-sentinel accept).
• Mutation-proven (skip each check → parity fails; <→<= over-strict → ties test fails; NA-set revert → over-reject reproduced).
• go build/vet/gofmt/go test/coverage-floor all green.

Follow-up (out of scope, flagged)

CheckSequenceRows/sequenceScanFrom (pre-existing) has the same NA-set bug family previewing SequenceGroupValidator (another plain-read_csv path): a sequence_id literally "none" is rejected by the CLI but accepted by the ingestor (over-reject), and "#NA" is under-rejected. Filed separately to keep this PR scoped.

Closes #218.

🤖 Generated with Claude Code

---

Note

Medium Risk
New grouped TSC validation mirrors pandas/ingestor semantics (NA tokens, dtype collapse, raw whitespace) and is parity-locked, but mistakes could still false-reject or miss issues; scope is limited to grouped TSC preflight only.

Overview
Adds local preflight for grouped time_series_classification so mid-sequence label changes and per-sequence unsorted timestamps fail the dry-run the same way the in-cluster di#359 validators do, instead of only after upload.

CheckTSCGroupIntegrity (single CSV pass via scanTSCRows) runs after null sequence_id checks in PreflightDataset when the category has the grouping trait. It previews one label per sequence (pandas groupby with null ids dropped, numeric label collapse when the column is all-numeric, raw cells for NA/object parity) and monotonic non-decreasing time within each sequence for schema-declared numeric time columns (ties and interleaved sequences allowed). TIMESTAMP/date ordering is intentionally not previewed to avoid over-rejecting vs pandas.

Grouped checks use pandasDefaultNA (STR_NA_VALUES), not the curated naSentinels used for schema-typed label diversity — fixing false rejects when #NA makes the label column numeric in-cluster.

Parity harness updates: tsc-label-flip and tsc-unsorted-timestamp now reject locally; new tsc-numeric-label-na-sentinel accept case; TestCheckTSCGroupIntegrity table tests the accept/reject boundaries.

^{Reviewed by Cursor Bugbot for commit fe46ff0. Bugbot is set up for automated code reviews on this repo. Configure here.}