Add imputation/bootstrap/SCHEMA.md

imputation/bootstrap/SCHEMA.md

@@ -0,0 +1,79 @@
+# Imputation bootstrap reference — schema
+
+The Track-2 bootstrap reference is two files:
+
+```
+imputation/bootstrap/draws.parquet      # ~107 MB zstd
+imputation/bootstrap/draws.meta.json    # provenance sidecar
+```
+
+## `draws.parquet`
+
+One row per `(method, scenario, split, channel, channel_type,
+subgroup_attr, subgroup_value, draw)`. Approximate size: **11,374,208
+rows** for the canonical 1000-draw / 16-method / 6-scenario / test-split
+run; `~107 MB` compressed (`zstd`).
+
+| column | type | description |
+|---|---|---|
+| `method` | string (dict) | method identifier (16 values: `locf`, `mean`, `mode`, `linear`, `temporal_mean`, `temporal_mode`, `personalized_mean`, `personalized_mode`, `personalized_temporal_mean`, `brits`, `dlinear`, `dlinear_weekly`, `fedformer`, `timesnet`, `lsm2`, `lsm2_weekly_sparse`) |
+| `scenario` | string (dict) | masking scenario (6): `random_noise`, `temporal_slice`, `signal_slice`, `sleep_gap`, `workout_gap`, `intensity_failure` |
+| `split` | string (dict) | data split — `test` |
+| `channel` | string (dict) | `ch_0`..`ch_18` (19 sensor channels) or `cat_collapsed:sleep` / `cat_collapsed:workouts` (collapsed-binary tasks) |
+| `channel_type` | string (dict) | `continuous` (`ch_0`–`ch_6`), `binary` (`ch_7`–`ch_18`), or `binary_collapsed` (the `cat_collapsed:*` tasks) |
+| `subgroup_attr` | string (dict) | `all` (global cell), `age_group`, or `sex` |
+| `subgroup_value` | string (dict) | `all` for the global cell; otherwise the subgroup level (age bucket, sex value, or `unknown`) |
+| `draw` | int32 | bootstrap-draw index in `[0, n_boot)` |
+| `E` | float32 | absolute error for this draw, user-macro reduced |
+| `R` | float32 | paired ratio vs baseline (`locf`) for this draw, geomean over users; `1.0` exactly for the baseline-vs-self row, `NaN` if undefined |
+| `rank` | float32 | cross-method rank for this draw, averaged over the resampled cohort |
+
+### Value semantics
+
+- **`E`** — per-draw user-macro error. Per channel: continuous = per-user MAE, then `nanmean` over resampled users; binary = `1 − AUC` per user, then `nanmean` over users; `binary_collapsed` = per-user `nanmean(1 − AUC)` over the category's channels, then `nanmean` over users.
+- **`R`** — per-draw paired ratio. Per (user, task) the ratio is `clip(E_method / E_baseline, 0.01, 100)`; per draw, `R = exp(nanmean(log(ratio)))` over resampled users. Binary `E` is `ε = 0.005` floored before forming the ratio so perfect-AUC users contribute finite values.
+- **`rank`** — per-draw cross-method rank. Per (user, task) the rank is `rankdata(E_methods, method="average", ascending=True)` (lower E → rank 1); per draw, `rank = nanmean` over resampled users.
+
+NaN appears where a task is structurally absent (e.g. binary tasks in the
+`EXCLUDE_BINARY_SCENARIOS` set: `sleep_gap`, `workout_gap`,
+`intensity_failure`).
+
+## `draws.meta.json`
+
+Side-car metadata, written next to `draws.parquet`. Schema:
+
+```jsonc
+{
+  "n_boot": 1000,
+  "seed": 42,
+  "splits": ["test"],
+  "scenarios": ["intensity_failure", "random_noise", "signal_slice",
+                "sleep_gap", "temporal_slice", "workout_gap"],
+  "methods": ["mean", "mode", "locf", ...],          // 16 entries
+  "method_dirs": {"locf": "/path/to/pairs", ...},    // source pair dirs (cluster-local)
+  "include_auc": true,
+  "include_fairness": true,
+  "age_bins": [18, 30, 40, 50, 60],
+  "exclude_unknown": false,
+  "elapsed_seconds": 3531.6,
+  "n_rows": 11374208,
+  "git_commit": "06e4cbda...",                       // code-repo commit at generation
+  "timestamp": "2026-06-19T03:24:21.860842Z",
+  "argv": [...]                                       // exact invocation
+}
+```
+
+## Conventions
+
+- Evaluated against the canonical split `sharable_users_seed42_2026` (`test`).
+- Track-2 baseline for skill / fairness: `locf`.
+- Format: single Parquet, dictionary-encoded categoricals, `float32` errors, `int32` draw index, `zstd` compression.
+- Resamples are paired across methods — same `boot_idx` matrix per split, so `R` and `rank` are valid per-draw comparisons.
+
+## Tracks
+
+| dir | track | status |
+|---|---|---|
+| `imputation/bootstrap/` | Track 2 — Imputation (above) | live |
+| `forecasting/bootstrap/` | Track 3 — Forecasting | added later |
+| `downstream/bootstrap/` | Track 1 — Outcome Prediction | added later |