[BUG] load_data_source validation gaps: non-numeric target passes, 1–2 row load blocked, file loader leaks internals #533

Severity: P1 (dtype) + P2 (row-count, error quality). Verified live 2026-08-09.

load_data_source validation runs but doesn't check the things fit actually needs, and several failure paths leak pandas internals. Three related findings:

NB-07 — non-numeric (object-dtype) target passes validation silently (P0-ish)

load_data_source(config={"type":"pandas",
  "data":{"date":["2024-01-01","2024-02-01","2024-03-01"], "value":["1","2","3"]},
  "time_column":"date","target_column":"value"})
-> {"success": true, "dtypes": {"value": "object"}, "validation": {"valid": true}}

inspect_data then returns categorical stats ({count, unique, top, freq}, no mean/min/max). It fails much later, inside fit. Same gap for object-dtype exogenous columns.
Fix: error, or at least warn, on a non-numeric target/exog dtype at load.

NB-06 — 1–2 row series cannot be loaded at all (P1)

load_data_source(... 1 row ...)
-> {"success": false, "error": "Need at least 3 dates to infer frequency", "error_type": "ValueError"}

This is a leaked pandas internal, and it's inconsistent: split_data happily creates 1-row handles, so 1-row data is representable — only loading is blocked.
Fix: load with freq=None + a warning, or a clear "at least 3 timestamped rows required to infer frequency" message.

NB-11 — file loader leaks parse_dates internals for a bad time_column (P2)

load_data_source(config={"type":"file","path":"...csv","time_column":"index", ...})
-> {"error": "Error reading CSV file: Missing column provided to 'parse_dates': 'index'"}

The pandas-source path gives a clean "Time column 'x' not found in data"; the file path should match it (and ideally list available columns).

(Catalogued as NB-07 / NB-06 / NB-11 in MCP_TEST_FINDINGS.md. Related to existing #313.)