Normalization: apportion a long unread gap across the months it covers

A counter delta is booked at the reading that closes it. That is right at the
reporting cadence -- a monthly series books December against the 1 January
reading, which is what the reference sheet does -- and wrong after an outage.
Observed on Solar 1: 78 days of generation arrived as one July row, leaving
June looking like the array was switched off.

An interval containing two or more complete calendar months is now divided
across them in proportion to elapsed time. The meter recorded a total, not a
shape, so every row a split produces is marked Estimated. The sum is exact: the
final segment absorbs the rounding remainder, so a split never creates or
destroys energy.

Counting whole months *contained* rather than boundaries *crossed* is what
makes the rule safe. A monthly series contains exactly one whole month per
interval and is untouched, so the golden fixtures keep measuring the normalizer
rather than the splitter; and a reading landing hours late cannot tip the rule
and hand the new month a sliver. GapSplittingIsInertOnFixturesTests asserts the
rule declines to fire on every reference interval, so this cannot drift into
the oracle unnoticed.

Not apportioned: swap and reset amounts (explicit corrections booked at their
event -- apportioning one would rewrite a number the operator supplied), a
rejected decrease, and a zero delta, which would otherwise fan out into rows
that say nothing.

Segments are stamped at their end, keeping the existing convention that a row
records the period ending at its timestamp -- so nothing shifts relative to how
unsplit intervals are already labelled.

Claude-Session: https://claude.ai/code/session_01V6joyergfvVLFEizH1hJLd

CLAUDE.md

@@ -72,7 +72,8 @@ sources (Tasmota/HA/MQTT/manual/CSV)
 
 **Invariants that shape everything:**
 
-- **Raw `reading` is immutable audit truth.** Everything derived (consumption, cost, balances, forecasts) is computed on top and must be reproducible. Never mutate readings to fix a derived number.
+- **Raw `reading` is immutable audit truth.** Everything derived (consumption, cost, balances, forecasts) is computed on top and must be reproducible. Never mutate readings to fix a derived number. Live ingestion recomputes the meter inline (`IngestionService.RenormalizeAsync`) — without it, polled readings never become consumption.
+- **Long gaps are apportioned, short ones are not** (`GapAttribution`, SDD §7.1). An interval containing ≥2 whole calendar months is split across those months, proportional to elapsed time, marked `Estimated`. A monthly series contains exactly one and is untouched — that's what keeps the golden fixtures reconciling. `GapSplittingIsInertOnFixturesTests` asserts the rule declines to fire on the reference data, so this can't silently drift.
 - **Dashboards and charts read aggregates only — never scan `reading`.** This is what makes 1000 meters × 50 years feasible (§5.5). Raw is kept for a bounded window (default 3y); `consumption` + aggregates are the long-term source of truth.
 - **`meter.mode` (measurement mode) is the central abstraction** for how raw readings become consumption (SDD §5.2): `cumulative_counter`, `generation_counter`, `runtime_counter` (Δhours × rate), `consumable_balance` (tank: deliveries − usage + forecast), `direct_delta`, `instant_rate`, `virtual` (expression over other meters). New ingestion/normalization logic dispatches on mode.
 - **Nothing domain-specific is hardcoded.** Energy types are data. Cost **categories are decoupled from energy types** (Heizung may be oil today, heat-pump tomorrow). PV self-consumption/savings/net are **virtual meters** with user-defined expressions, not special-cased code. Tariffs are time-ranged (price history), scoped global / per-type / per-meter.