# Handoff — 2026-09-26 (before HPC maintenance)

**HPC maintenance / shutdown:** Monday **2026-09-28, 00:00 UTC = 07:00 local (UTC+7)**.

## In-flight job: `pl_hi` bulk download
- Command: `python3 scripts/05_download_reforecasts.py --group pl_hi --dates all`
- Started 2026-09-26 ~14:54 local, running in the background (was PID 1543).
- Progress at handoff (~16:xx local): **20/104 `pl_hi` dates on disk**, last started #19 (mv2023-10-26).
- Rate ~10 dates/hour, no ECDS rejections. Projected finish **~01:00 local Sun 2026-09-27** — ample margin before the Monday outage.
- **Resumable**: skip-existing + retry/backoff. If interrupted (incl. by the shutdown), just re-run the same command after maintenance; it continues from the missing dates.
- Output log (this session only): `/tmp/claude-504/-home-tanpv-Progs-claude-FM/.../tasks/b253nc0tc.output`

### To check status later
```bash
ls data/raw/ecmwf_s2s_1p5deg/ | grep -c s2s_cf_pl_hi_ls    # want 104
pgrep -af 05_download_reforecasts                           # is it still running?
```

## Data state (reforecast, mv2023 × hy2003-2022, native 1.5°)
- `sfc_tx_loc`  — **104/104** ✓ (daily-max Tx source, mx2t6, 6-hourly)
- `sfc_dyn_loc` — **104/104** ✓ (2t, msl, 10u, 10v, daily)
- `pl_hi_ls`    — **20/104** (in progress; target 104) — T, Z, U, V, q @ 850/700/500/200 hPa, LARGE domain
- Full predictor set = all three groups at 104 dates. See `docs/RETRIEVAL_PLAN.md`.

## Git
- Checkpoint commit `28ac2da` on branch **`checkpoint/pre-hpc-maint-2026-09-28`** (working tree snapshot: B1 prototype, this plan, grib_preprocess, download script, config/decision-log edits).

## Other-project jobs on this account (DO NOT TOUCH)
- `Dr_GCM/.../download_ecmwf_baseline` (separate project) is running under the same user; may share the ECDS queue. Left untouched per project rules.

## Next work after `pl_hi` completes (local compute; HPC-up but no network needed)
1. `scripts/08` assembly → per-variable cube `(model_version_date[104], hindcast_year[20], lead[46], lat, lon)` on native grids (predictors NOT regridded to 0.25° pre-encoder).
2. Then the ML prototype on the assembled cube.

## Verified this session
- B1 prototype ran end-to-end (exit 0): Tx MAE 1.60→2.62 °C (LT0→LT45), EHD POD 0.51→0.06, over 586 cells. Artifacts in `outputs/B1_PROTOTYPE.html`, `outputs/metrics/b1_prototype_by_lead.csv`, `outputs/figures/b1_prototype_leadcurves.png`.