test(parity): value-level label parity — catch accept/accept-with-divergent-data (#1009 P0)

[LukasWodka]

Summary

The validator-parity harness pinned only accept/reject verdicts. So the one failure mode RFC-0002 leans hardest on — the CLI previews "N labeled rows, K classes" while the ingestor actually writes null/fewer — sailed through every gate. That's the data-ingestors #340 class (a case-/whitespace-mismatched label passes both verdicts, then reads NULL in-cluster). This adds a value-level layer on top of the verdict pins: the resolved label header + row count + class set the real ingestor read path produces must equal the Go preview's.

This is the P0 "value-level parity" keystone of #1009 — "the only thing that catches accept/accept-with-divergent-data."

What changed

• gen-validator-goldens.py — for cases flagged value_parity (and accepted), drives the REAL ingestor read path (CSVIngestor.read_data + the chore: back-merge main → develop (adopt cobra 1.10.2 + reconcile drift) #340 label resolution + RecordProcessor) and records {resolved_label, row_count, classes}. Fails loudly if the target ingestor predates chore: back-merge main → develop (adopt cobra 1.10.2 + reconcile drift) #340 (won't pin the bug).
• preflight.go — factored CheckLabelDiversity's read/resolve/NA/collapse loop into readLabelColumnValues (one scan → resolved label, class set, row count) and exposed ReadLabelValues. CheckLabelDiversity now derives its verdict from that single source — no forked read loop; also gives RFC-0002's "check your data" preview a real value read to show.
• parity_golden_test.go — for value_parity cases, asserts the Go preview's resolved label + row count + class set equal the ingestor golden.
• sync-validator-goldens.sh --check — now compares values too (they carry no fixture paths), so a value-only drift can't slip past the verdict-only check.
• cases.json — flags the 9 accept/accept string-label cases as value_parity.

Proven to bite

Mutating imgc-label-case's golden to the pre-#340 state (unresolved label, empty classes) fails the test:

resolved label: Go preview = "Label", ingestor golden = "label"
(the read paths resolve the label column differently — #340 class)
class set: Go preview = [cat dog], ingestor golden = []

Regenerating against #340 restores green. imgc-label-case (header " Label " vs config label) is exactly the `Label`/label read-path case #1009 P0 names — value parity turns it into a live proof of #340.

Scope & dependency

• String-read labels only (image/text always; tabular string-family). Numeric labels stay verdict-only — cross-language float formatting ("1.0" vs "1") makes value comparison unreliable, and the numeric collapse is already verdict-pinned.
• Depends on data-ingestors #340: the value goldens are generated against it. Once chore: back-merge main → develop (adopt cobra 1.10.2 + reconcile drift) #340 lands on data-ingestors develop, a regen is a no-op; the CI drift job (below) must pin a data-ingestors SHA that includes chore: back-merge main → develop (adopt cobra 1.10.2 + reconcile drift) #340.

Type

Test hardening · cli · RFC-0002 foundation. Part ofbackend#1009 (P0 value-level parity) — not the whole ticket.

Remaining #1009: money-path seam (submitFn/portForwardFn/mintTokenFn + exit-code/reclaim matrix), CI drift wiring (pin the schema-drift + goldens SHA, per-package coverage floor), cross-repo taxonomy contract test, one real content-compared ingest e2e.

---

Note

Low Risk
Test-hardening and a refactor of label reading that keeps diversity checks on the same scan; production preflight behavior should be equivalent with clearer shared semantics.

Overview
Extends the validator-parity harness beyond accept/reject so the Go CLI preview cannot agree with the ingestor on verdict while reading different label data (data-ingestors #340).

Preflight refactors label diversity into a shared readLabelColumnValues scan that returns resolved header, row count, and class set via ReadLabelValues / LabelReadValues; CheckLabelDiversity now derives its verdict from that same read.

Harness & goldens:cases.json marks nine accept cases with value_parity. gen-validator-goldens.py records ingestor value goldens from the real CSVIngestor read path (requires #340_resolve_label_column). parity_golden_test.go asserts Go values match those goldens. sync-validator-goldens.sh --check compares verdicts and values so value-only drift fails CI.

Numeric-label cases stay verdict-only; string-read paths are what value parity pins.

^{Reviewed by Cursor Bugbot for commit 00f728e. Bugbot is set up for automated code reviews on this repo. Configure here.}

[saadqbal]

bugbot run

[cursor]

Bugbot couldn't run

Bugbot is not enabled for your user on this team.

Ask your team administrator to increase your team's hard limit for Bugbot seats or add you to the allowlist in the Cursor dashboard.

[LukasWodka]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 00f728e. Configure here.}

[saadqbal]

Nice, careful PR 👍 The single-scan refactor of CheckLabelDiversity is behavior-preserving (all 25 parity cases still pass) and the pre-#340 mutation test is a solid proof it bites. One non-blocking thought: read_label_values value-parity leans on infer_schema typing everything VARCHAR — worth a comment noting numeric-looking labels need an explicit schema, though that asymmetry predates this PR.