feat(push): infer entirely-empty columns as nullable FLOAT + warn

[LukasWodka]

Summary

InferSchema typed an entirely-empty column (no non-empty value in the sample) as VARCHAR(255). That column then fails ingest: data-ingestors' string validator rejects the all-NULL column. Sparse panels hit this on every empty column.

This types such columns as a nullable FLOAT instead (the FLOAT validator accepts NULLs, so an all-empty column ingests cleanly as NULLs), and surfaces them so dataset push warns the user.

Why

A real biomarker panel (proteomics matrix) — an analyte measured in zero samples is an all-empty column:

patient_id,Sex,...,P02751-1|FN1,P05109|S100A8
PAT000001,female,...,,18.6198268 # FN1 empty for every patient

FN1 (empty everywhere) inferred VARCHAR(255) → ingest failed with Column 'P08254|MMP3' contains 207 non-string values (207 patients, none measured). Sparse panels always have unmeasured analytes, so this bit every empty column.

What

• InferSchema: a column with no non-empty sampled value → FLOAT (was VARCHAR(255)), and is returned in a new empty []string so the caller can warn. The numeric path is unchanged (all-int → INT, all-numeric → FLOAT, otherwise VARCHAR).
• dataset push prints a warning naming the empty columns (typed FLOAT/nullable; override with --schema if a column is really text).
• Updated the EmptyColumn test (now expects FLOAT + checks the empty return) and the two other InferSchema callers for the new signature.

gofmt + go vet + go test ./internal/push/ ./internal/cli/ all pass.

Pairs with

data-ingestors#167 — fixes the validator so a NULL is never miscounted as "non-string." This CLI change is the primary unblock (empty col → FLOAT → passes); #167 is the safety net for genuinely-text columns that have blanks. Either alone helps; together they close the sparse-tabular trap.

Note / open question for review

I changed the documented empty→VARCHAR(255) fallback to empty→FLOAT. Rationale: a tabular feature column is numeric far more often than text, and FLOAT-nullable ingests where all-NULL VARCHAR doesn't (and it also covers the "empty in the first 5k sampled rows, numeric later" edge). The warning + --schema override cover the rare genuinely-text-empty column. Happy to flip to "warn-only, keep VARCHAR, rely on #167" if you'd rather not change inferred types — your call.

🤖 Generated with Claude Code

[LukasWodka]

👋 Heads-up — Code review queue is at 18 / 8

Above the WIP limit. The team convention is to review existing PRs before opening new work.

Open PRs currently in Code review (oldest first):

• averaging-service#103 — fix(averaging): correctness foundations — DP, shape guard, sklearn weighting, bounded ensembles (dataset push: tabular pre-flight silently collapses duplicate header columns and accepts zero-data-row CSVs #73) · author: @LukasWodka · reviewer: @shujaatTracebloc
• backend#732 — feat(experiments): add per-experiment aggregation_strategy field (#730) · author: @LukasWodka · reviewer: @saqlainsyed007, @shujaatTracebloc
• backend#735 — fix(flops): atomic flops_used + fire stop once on threshold crossing (#733) · author: @saadqbal · reviewer: @aptracebloc
• cli#64 — release: promote develop → main (toward v0.3.1 — install.sh PATH fix) · author: @saadqbal · no reviewer assigned
• cli#65 — fix(install.sh): tighten PATH idempotency + persist user-local PREFIX (Bugbot on fix(install.sh): persist PATH to the shell rc (parity with install.ps1) #61) · author: @saadqbal · no reviewer assigned
• client#221 — release: promote develop → main (installer hardening — OOM gate, apt/needrestart fixes, drift CI) · author: @saadqbal · no reviewer assigned
• data-ingestors#167 — fix(validators): VARCHAR/CHAR/TEXT must tolerate NULLs · author: @LukasWodka · reviewer: @saadqbal
• model-zoo#87 — feat(tabular): stability-selection sklearn models (LASSO + RF) · author: @LukasWodka · no reviewer assigned
• tracebloc-client#355 — fix(object_detection): scale RCNN targets to pixel space · author: @shujaatTracebloc · reviewer: @divyasinghds
• tracebloc-client#356 — Promote develop → staging: graceful exit on backend-stopped experiment (doctor/home: stale "Online" false-green + duplicated hint literal (minor) #353/fix(doctor): stop the cluster-down cascade of misleading remedies #354) · author: @saadqbal · no reviewer assigned

Pull from review before opening new work. (This is a nudge from the kanban WIP check, not a block.)