fix(ha-ws): correct cumulative sum handling on incremental import

[jrozelle]

Summary

The HA WebSocket exporter (recorder/import_statistics) has 3 distinct bugs that together break the Energy dashboard for users running incremental imports. They compound: each one masks the next.

This PR fixes all three with a single localized change to home_assistant.py.

Bug 1: HA timestamps are milliseconds, not seconds

get_last_statistic_dates parses the start field of recorder/statistics_during_period responses with datetime.fromtimestamp(start_ts) — but HA core 2023+ returns these timestamps as a numeric epoch in milliseconds, not seconds.

Effect: OverflowError: year 58244 is out of range. The function falls through to the bare except, returns success: False, and the caller silently falls back to a full re-import on every scheduled run.

Fix: auto-detect ms vs s by magnitude (start_ts > 1e12 → ms). Safe until ~year 33658 in seconds.

Bug 2: Cumulative sums restart at 0 in incremental mode

_get_consumption_statistics_by_tariff, _get_cost_statistics_by_tariff and _get_production_statistics initialize their cumulative accumulators to 0.0 unconditionally. In full mode this is correct. In incremental mode, the new sums clobber the existing monotonic series in HA — the Energy dashboard sees a giant negative delta on the boundary hour (e.g. -11705 kWh on TEMPO blue_hp at the season transition) and stops displaying anything sensible.

Fix:

• get_last_statistic_dates now also returns last_sums (extracted from the same statistics_during_period response, no extra round-trip).
• A new helper _extract_per_pdl_state splits the flat {statistic_id: sum/date} maps into per-category buckets needed by the 3 export functions.
• The 3 builders accept initial_sums (or initial_sum for production) and seed their cumulative accumulators with those values instead of 0.0.

Bug 3: Per-tariff records re-processed and double-counted

After fixing bug 2, the cumulative continuity worked but each new MED run shifted the series upward by the consumption delta between the global oldest_date and each tariff's own last_date.

Root cause: incremental mode uses ONE global since_date (the oldest across all tariffs) for the SQL fetch. But each individual tariff has its own newer last_date — and the previous code re-iterated all fetched records, re-importing entries that were already in HA and overwriting their sums with shifted-up values. After enough runs, totals are 2-3x the real value.

Fix: pass last_dates_by_tariff (and last_date for production) to the builders, and skip records whose timestamp is <= last_date for that tariff. The SQL fetch stays wide (global oldest) so every tariff still sees its new records, but per-tariff filtering inside the loop prevents re-processing.

Cost statistics inherit the filter for free since they're derived from the (already filtered) consumption_by_tariff.

Effect

In incremental mode, MED now correctly:

• Reads existing sums from HA and uses them as cumulative starting points
• Skips records that are already exported (per tariff)
• Only ADDS strictly new records on top of the existing series
• The sum stays monotonic, no compensation event in HA Energy

Test plan

Verified locally on a TEMPO production setup running for 26 months:

• After the fix, no year 58244 error in MED logs (was happening every 30 min)
• After the fix, MED reports Found last dates for 12/13 statistics and runs in incremental mode (not full-import on loop)
• After the fix, sums stay monotonic across all 4 season-transition days (2026-03-31, 2026-04-01, 2026-04-03)
• After the fix, the 12-month rolling total in HA Energy matches actual consumption (~10 MWh/year vs ~21 MWh that the bugs were producing)
• After the fix, repeated incremental runs do not shift existing entries (verified by querying the recorder DB directly between runs)
• Compose-build the patched container, confirm Python startup is clean

Note: this PR only contains the code fix. Existing users whose recorder DB was already corrupted by the bugs will need a one-shot data repair (purge med:* stats from HA recorder and let MED full-reimport). A separate runbook can be added if needed.