fix(cso): normalize incidentDate/submittedDate to a uniform string on write

[nesanders]

Problem

The 2026-08-22 dispatched Weekly Update run (32592642457) got past the lobbying-fetch timeout (fixed by #111/#112 — confirmed, no longer an issue) but then crashed in "Generate dashboard charts":

ValueError: time data "2026-08-08" doesn't match format "%Y-%m-%d %H:%M:%S"

in EEA_DP_CSO_map.py's load_data_cso(), the first parse of MAEEADP_CSO.incidentDate read back from AMEND.db.

Root cause

Reproduced locally against the live EEA CSOAPI. incidentDate is always midnight in the API — time-of-day lives in a separate incidentTime field. This week's incremental batch of new incidents happened to be entirely midnight values; pandas' to_csv formats that batch as bare "YYYY-MM-DD" while older rows in the same file keep "YYYY-MM-DD HH:MM:SS" — producing a genuinely mixed-format CSV column that a plain pd.to_datetime() re-parse can't handle.

A second, related gap: get_data() parsed incidentDate on load but never submittedDate, so pd.concat([existing, new_df]) silently produced a mixed object column (raw CSV strings next to real Timestamps) for submittedDate too — already present in the committed CSV, just never triggered because nothing downstream re-parses that column directly.

Fix

• get_eea_dp_cso.py get_data(): also parse existing['submittedDate'] on load (matching incidentDate), so both columns are uniformly typed before concat.
• get_eea_dp_cso.py write_data(): explicitly normalize both date columns to a canonical "%Y-%m-%d %H:%M:%S" string (format='mixed' before formatting) before writing, instead of relying on pandas' to_csv default formatting.
• EEA_DP_CSO_map.py load_data_cso(): format='mixed' on the read-side parse as a defensive fallback, so a future anomaly degrades gracefully instead of crashing the whole dashboard build.

Verification

• incidentDate values preserved exactly (0 mismatches vs. previously-committed data).
• submittedDate differs only by sub-second truncation (max 997ms, zero full-second+ diffs) — intentional; that field is never used for precision-sensitive analysis.
• The exact crash-reproducing call (pd.to_datetime(), no format, on the written column) now succeeds.
• Includes the freshly re-fetched, now-uniform CSO data (through 2026-08-22) since the crashed run never reached the commit step.

🤖 Generated with Claude Code