daily-fix: fail-fast SFT prompt/completion type assert

Overview / Motivation

Auto-filed by the 2026-07-18 /daily Step C parked-candidate sweep from a prose follow-up raised on task #1508 (emitting agent: Methodology critic). NOTE: this is an EXPERIMENT/LIBRARY-code fix (src/explore_persona_space/train/ is outside the workflow-fix surface) — filed as a wf_fix: false route-2 item per the /daily route-2 variant; no wf-fix tags, no Provenance injection.

Goal

Add a deterministic fail-fast prompt/completion type-homogeneity assert in the project SFT data path (src/explore_persona_space/train/sft.py / train_lora): validate that BOTH keys agree in type across the dataset before SFTTrainer init, as the code-level backstop to the #1508 doc entry.

Workflow gap

• Bug observed: a mixed-type prompt/completion dataset enters the SFT data path undetected until deep inside training (#1508's incident); the existing isinstance(prompt, list) checks in sft.py sit in the marker-probe and KL-aux side paths and SILENTLY SKIP non-conforming rows (return None / continue) rather than failing fast.
• Why it is a workflow gap: (experiment-code gap, not workflow-surface) the project's fail-fast rule ("no silent defaults; the crash IS the signal") is violated by the silent-skip shape; a homogeneity assert before SFTTrainer init converts a late confusing failure into an immediate, attributable one.
• Confidence (emitter): medium
• verified-at-filing: grep -n 'isinstance' src/explore_persona_space/train/sft.py → type checks at L833 + L1129; context READ (L820-845, L1120-1140): both are silent-skip guards in the marker-slot probe and KL-aux batch builder, NOT a fail-fast homogeneity validation of the training dataset before SFTTrainer init — the proposed assert is not landed; grep -n 'homogene' src/explore_persona_space/train/sft.py → 0 hits; git log --oneline --since='7 days ago' -- src/explore_persona_space/train/sft.py → 0 commits (2026-07-19)

Proposed change (candidate diff sketch — refine in planning)

(none — synthesized from prose follow-up)

Sketch for the planner: in the sft.py dataset-load/validation path (before SFTTrainer construction in train_lora), scan the loaded rows once and raise a ValueError naming the first offending row index when prompt / completion types are non-homogeneous (str-vs-list mixed) or the two keys disagree in convention; cover with a unit test on a deliberately mixed JSONL.

Scope / surfaces

• Primary target: src/explore_persona_space/train/sft.py (train_lora data path)
• Experiment/library code — the standard /issue code-change pipeline (implementer + code-reviewer + tests) applies; this is NOT a workflow-surface edit.

Constraints / invariants

• Fail fast, never silently skip rows in the TRAINING path; the existing probe/KL-aux silent-skip guards may stay (they are diagnostic side paths) but the plan states the boundary explicitly.
• No behavior change for currently-valid homogeneous datasets; ruff + the sft.py pinning tests pass.

Provenance

• source task: #1508 (prose follow-up, methodology critic, 2026-07-18T13:30:35Z)
• fingerprint (informational; wf_fix false — no fp-dedup): 8db8ed7096eb

Verbatim surfaced prose (task #1508 events.jsonl): "Methodology critic surfaced-prose follow-up — add a deterministic fail-fast prompt/completion type-homogeneity assert in the project SFT data path (src/explore_persona_space/train/ sft.py / train_lora, validate both keys agree before SFTTrainer init) as a code-level backstop to the #1508 doc entry. source: prose-followup; confidence: medium; related_task: #1508."