fix(push): fail closed on labels.csv read errors in label + sequence preflight

[saadqbal]

What

Two more preflight mirror-checks swallowed a non-EOF csv.Reader.Read error with continue — the same fail-open class as #221. Follow-up Bugbot findings on the promote PR #219.

• readLabelColumnValues (backs CheckLabelDiversity + ReadLabelValues): a mid-read failure returned a partial class set with Found=true. An under-counted column could false-reject good data or let a bad file pass. Now returns an error → the diversity gate fails closed; the value preview (ReadLabelValues) degrades to Found=false.
• CheckSequenceRows: a mid-read failure skipped the unread tail, so null/missing sequence_ids there never surfaced locally. Now fails closed.

CrossCheckLabels and the tabular schema scan already abort on the same error — this makes the whole internal/push package uniform.

Trigger (same correction as #221)

Not malformed CSV content: openCSVReader sets LazyQuotes + FieldsPerRecord=-1, so every bad shape parses like pandas. The branch is only reachable via a genuine I/O read error.

Test

Each scan is split into a reader-based core (labelColumnValuesFrom / sequenceScanFrom) so the branch is exercised with an injected fail-after-header reader (reusing failAfterReader from #221). Parsing is byte-for-byte unchanged; the value parity harness passes, proving the relocation is faithful.

Scope

Bugbot drive-bys on #219; roll up under release ticket #220 (no separate kanban item). Targets develop → joins the v0.8.0 payload with #221.

🤖 Generated with Claude Code

---

Note

Low Risk
Localized preflight error-handling only; happy-path parsing is unchanged and parity harness coverage is noted in the PR.

Overview
Label and sequence CSV preflight no longer treat a mid-file csv.Reader I/O error as a skippable row. That path used to continue and could return a partial class set or pass a scan whose tail never ran—matching the fail-open bug fixed in #221 for text/CrossCheckLabels.

readLabelColumnValues now returns an error on mid-read failure; CheckLabelDiversity propagates it and fails closed. ReadLabelValues still ignores that error and reports Found=false so previews do not show wrong counts. The scan logic moves into labelColumnValuesFrom on an injected reader for tests.

CheckSequenceRows does the same for sequence ids via sequenceScanFrom: read errors surface as wrapped failures instead of hiding null/missing ids in unread rows.

New unit tests use the existing failAfterReader helper to assert both paths abort with the underlying error and do not mix read failures with domain null-id errors.

^{Reviewed by Cursor Bugbot for commit fc97b90. Bugbot is set up for automated code reviews on this repo. Configure here.}

[saadqbal]

Added the root-cause fix on top of the fail-closed change: openCSVReader set FieldsPerRecord=-1 but not LazyQuotes, so a bare/unescaped quote (which pandas tolerates) made Read error. That meant every preflight mirror-check false-rejected a CSV the ingestor accepts — and made the fail-closed change reachable via a bare-quote row (fail-open → false-reject). Now LazyQuotes is set centrally (+ in InferSchema's own reader), so the readers are pandas-faithful and the remaining read error is genuinely I/O-only. Full parity harness still passes + added a bare-quote guard test. This is the real fix behind the read-error findings Bugbot kept surfacing.