Phase 7: precipitation, S2F disaggregator, CBRFC LID helper, polish

[tmart234]

The model didn't see precipitation at all -- not observed (encoder) and not
forecast (decoder). For a 14-day Colorado snowmelt forecast that's a real
omission; SWE and antecedent soil moisture covered "water in the system"
but not recent or upcoming rain. Add a precipitation feature wired
end-to-end:

• get_noaa.py now fetches GHCND PRCP alongside TMIN/TMAX and converts
  tenths-of-mm to mm so units match Open-Meteo's precipitation_sum.
  Adds a realized-coverage gate (fix for the long-standing TODO at the
  top of the module): walks past stations whose actual returned data is
  too NaN-heavy on TMIN/TMAX, not just the lifetime metadata coverage.
• get_forecast.py adds precipitation_sum to the Open-Meteo request and
  gracefully falls back to zeros if an older deployment omits it.
• combine_data.py threads precipitation through merge_dataframes with a
  short (2-day) interior interpolation limit -- precip is spiky and
  smearing a missed storm across surrounding dry days is misleading --
  and 0-fills anything still missing (modal value for most basins).
• normalize_data.py registers precipitation as optional + log1p-scaled
  (heavy-tailed, same rationale as flow).
• windowing.py adds precipitation to both ENCODER_FEATURES (observed)
  and DECODER_FEATURES (forecast QPF) -- the one auxiliary that has a
  skillful 14-day forecast available.

Other items in the same sweep:

• get_s2f.py: ships disaggregate_seasonal_to_daily() so when someone
  wires the USBR archive fetch, the baseline lights up immediately
  instead of needing the math written then too.
• data/find_lid.py: maintainer helper that queries the NWS NWPS gauge
  index near a USGS/DWR site and prints candidate AHPS LIDs, so the
  CBRFC LID map gets populated with real values instead of guesses.
• get_poly.py: HUC-aware simplification tolerance (HUC4 polygons have
  ~10x the points of HUC8; same fixed tolerance was over/under-doing it
  depending on the level).
• export_mobile.py: optional dynamic-range int8 TFLite output, gated on
  a parity check vs the float32 Keras model. Ships only when max abs
  diff < 0.05 (scaled space); manifest records the parity number.
• VegDRI removed (two never-wired modules + their xfailed test). Two
  years of dead code, no integration path that didn't require a new
  USGS time-series scraper from scratch.
• docs/INFERENCE.md: documents precipitation handling, int8 artifact,
  and the gh workflow run one-liner for the first release after merge.

https://claude.ai/code/session_01XfhRQmztLSqmz6qeLSi9kw