fix(weather): resolve GFS precip twin crash + expose cloud_cover_pct, visibility_m, cloud_ceiling_m (#63)

[zax0rz]

Summary

Fixes #63. Resolves a latent GribIntegrityError crash that affects every
forecast_nwp(station, "gfs") call at the default fxx=1, and adds three
new NWP forecast columns: cloud_cover_pct, visibility_m, and
cloud_ceiling_m for both HRRR and GFS models.

The latent GFS precip bug

GFS GRIB2 files contain duplicate APCP:surface records at fxx >= 1
(two identical accumulation windows with different record numbers). The SDK's
ambiguity guard in _extract_records() raises GribIntegrityError on any
duplicate, completely aborting the fetch. Both mirrors carry the same twins, so
mirror fallback doesn't recover. At fxx=0 the bug is masked because GFS omits
APCP entirely at the analysis hour.

This was invisible because CI has no live GFS test — only HRRR.

Fix: _pick_record heuristic

Replaces the fatal GribIntegrityError with a deterministic picker that:

1. Prefers instantaneous records over window-aggregated (ave/acc/max/min)
2. Breaks ties by lowest record_no

This simultaneously fixes the APCP crash and unblocks cloud_cover_pct on GFS,
where (TCDC, entire atmosphere) also produces two records (instantaneous vs
time-averaged). HRRR is unaffected — all fields are single-record.

New columns

Column	GRIB2 Key	Level	cfgrib short-name	Units
cloud_cover_pct	TCDC	entire atmosphere	tcc	%
visibility_m	VIS	surface	vis	m
cloud_ceiling_m	HGT	cloud ceiling	gh	m

Naming uses cloud_ceiling_m (not ceiling_m) for cross-source consistency
with IEM adapters.

Supporting evidence

Academic and operational research in .briefs/ demonstrates why these fields
matter for prediction-market temperature modeling:

• Cloud cover modulates DTR by 50%+ — overcast vs clear-sky shifts daily
  highs by 7–11°F and lows by 8–10°F. That's the exact margin around Kalshi
  strike prices.
• NWS MOS uses cloud cover as a primary predictor in its temperature
  regression equations. The SDK was missing what MOS considers essential.
• BC-Unet (NOAA's own ML bias correction for GFS) ingests total cloud
  cover as a core 2D input field for image-to-image temperature correction.
• HRRRv4 systematically underestimates cloud fraction — exposing TCDC
  enables quants to detect and correct for this known radiative bias.

See .briefs/cloud-cover-deep-research.md (64 cited sources) and
.briefs/github-issue-63-nwp-fields-review.md for full analysis.

Changes

Bug fix

• forecast_nwp.py: Add _pick_record() disambiguation helper
• forecast_nwp.py: Replace raise GribIntegrityError with warning + picker

New fields

• schemas/forecast_nwp.py (core): Register 3 nullable float64 columns
• _nwp_grids/hrrr.py, _nwp_grids/gfs.py: Add TCDC, VIS, HGT mappings
• forecast_nwp.py: Add cfgrib short-name lookups (tcc, vis, gh)
• forecast_nwp.py: Register in nullable_numeric_cols and _empty_dataframe
• qc/rules_nwp.py: Add physics-bounds QC rules (cloud 0–100%, vis/ceiling ≥ 0)
• schemas/json/schema.forecast_nwp.v1.json: Regenerated

Tests

• TestDisambiguationHeuristics: synthetic duplicate .idx fixtures, window vs
  instantaneous preference, record_no tiebreak
• test_qc_rules_nwp.py: updated base rule count (7 → 10)
• test_forecast_nwp_multi_cycle.py: mock row structures updated

Test plan

• 1459 passed, 1 skipped, 23 deselected (live excluded) — uv run pytest packages/weather/tests -m "not live"
• 2938 passed, 4 skipped, 48 deselected — full suite uv run pytest -m "not live"
• Ruff check + format clean
• Committed on branch fix/63-nwp-cloud-cover-precip from upstream/main

Cross-SDK Sync (see .planning/CROSS-SDK-SYNC.md)

python_only: true — external contributor PR; TS parity ticket will be filed
by a maintainer after merge. The schema column additions and .idx
disambiguation heuristic both need TS ports (documented in PR body and
.briefs/github-issue-63-nwp-fields-review.md §6).

[helloiamvu]

Codex review (gpt-5.5, medium reasoning) — 1 blocking finding (P2):

• [P2] Exported NWP schema not wired into TS codegen — scripts/export_schemas.py:97. schema.forecast_nwp.v1 is added to the exported Group A schemas, but the TS generator's schema list is hard-coded without schema.forecast_nwp.v1.json, so pnpm codegen won't emit a ForecastNwpV1 type/validator and TS consumers stay unaware of the new public columns. Per the CLAUDE.md dual-SDK rule, update the TS codegen schema list and regenerate the generated TS files alongside this export (or file the parity ticket the issue's TS-Parity section anticipates).

CI is green and the Python disambiguation logic + GFS-precip fix look solid; this is the one issue to resolve before merge. (Reviewed in the 2026-06-06 autonomous triage run — feature-scope, recommend a 1.6.0 minor.)