docs(tau): final five-arm tau=9 retrain table + slice rationale (follow-up to #33)

.claude/skills/ddrs-dev/references/research-status.md

@@ -279,8 +279,15 @@ was invalidated by a stale binary and the manifest did not reveal it.
   **Convention change SHIPPED (2026-08-08, findings §5i):** tau is now
   signed-at-zero hours of advance (`[tau : -(24-tau)]`, day i ↔ obs day i,
   default 9 ≡ old 20); old scale = new + 11. Old checkpoints trained at
-  old-3 ≡ new −8. tau=9 CPU retrain across all five stores launched same
-  day (`scripts/run_tau9_source_trains.sh`).
+  old-3 ≡ new −8. **Retrain COMPLETE (2026-08-10, all five stores, 30
+  epochs, CPU, 1,841 gauges, full test window): aorc2f distributed
+  0.620 → 0.706 median NSE (+0.086 from timing alone; now beats the
+  summed-q' baseline 0.642 by +0.064), UH retro 0.707, daily LSTM 0.578,
+  hourly LSTM 0.564, aorc2f lumped 0.483 (WORSE, as pre-registered — its
+  optimum is ≈ −8; the negative control behaved).** Merged to master in
+  PR #33. Gotcha: a train-and-test resume from a final checkpoint cannot
+  re-enter Phase 2 (needs checkpoints from its own Phase 1) — finish a
+  killed eval with the legacy eval binary instead.
   **Gamma-UH params pulled (§5h):** the distributed aorc2f store's q' was
   exported (2026-07-29, water_loss) with each divide routed through its own
   learned gamma UH; `scripts/dump_gamma_uh_params.py` (water_loss venv)

docs/2026-08-06-tau-sweep-pilot-findings.md

@@ -509,8 +509,8 @@ Caveat noted at launch: the aorc2f_lumped store's measured optimum is
 new-convention ≈ −8 (§5e), so tau=9 is expected to HURT that arm; it runs
 anyway as the consistency control.
 
-**Retrain results (2026-08-09, full test window 1995-10..2010-09, 1,841
-gauges, 30 epochs each):**
+**Retrain results — FINAL, all five arms (2026-08-09/10, full test window
+1995-10..2010-09, 1,841 gauges, 30 epochs each, CPU):**
 
 | arm | median NSE | median KGE |
 |---|---|---|
@@ -519,17 +519,26 @@ gauges, 30 epochs each):**
 | **tau=9 aorc2f distributed** | **0.706** | **0.730** |
 | **tau=9 UH retrospective** | **0.707** | **0.738** |
 | tau=9 daily LSTM | 0.578 | 0.616 |
+| tau=9 hourly LSTM (native hourly) | 0.564 | 0.500 |
+| tau=9 aorc2f lumped (consistency control) | 0.483 | 0.504 |
 
 The timing fix alone is worth **+0.086 median NSE** on the flagship arm
 and flips routing from LOSING to the summed-q' baseline (0.620 vs 0.642)
-to beating it by +0.064. Run IDs: `2026-08-09T{03-05-54,09-30-39,12-05-08}Z-
-train-and-test`. Operational note: the hourly_lstm arm finished training
-(run `2026-08-09T14-55-05Z`, 60 checkpoints) but its eval phase was killed
-with the parent driver at chunk 49/366; a train-and-test resume from the
-final checkpoint cannot re-enter Phase 2 (it requires checkpoints written
-by its own Phase 1 — `tau9_train_hourly_lstm_evalresume.yaml` records the
-failed attempt), so the arm completes via the legacy eval binary
-(`scripts/run_tau9_hourly_eval_chain.sh`), chained behind the
+to beating it by +0.064. Every arm ordered exactly as the eval-only
+sweeps predicted (§5e), including the negative control: the lumped store
+(measured optimum ≈ −8 on the new scale) got WORSE at tau=9 (0.510
+eval-only at old tau → 0.483 trained at tau=9) — the one arm the theory
+said the fix should hurt, hurt. Run IDs:
+`2026-08-09T{03-05-54,09-30-39,12-05-08,14-55-05,22-13-43}Z-train-and-test`.
+
+Operational note: the hourly_lstm arm finished training (run
+`2026-08-09T14-55-05Z`, 60 checkpoints) but its eval phase was killed with
+the parent driver at chunk 49/366; a train-and-test resume from the final
+checkpoint cannot re-enter Phase 2 (it requires checkpoints written by its
+own Phase 1 — `tau9_train_hourly_lstm_evalresume.yaml` records the failed
+attempt), so its numbers above come from the legacy eval binary against
+the epoch-30 checkpoint (`scripts/run_tau9_hourly_eval_chain.sh`,
+`output/tau_sweep/train9_hourly_lstm_eval.{log,zarr}`), chained behind the
 aorc2f_lumped run (`scripts/run_tau9_remaining.sh`).
 
 ## 6. Raw output

src/training/loss.rs

@@ -29,6 +29,19 @@ use crate::config::{LossConfig, LossKind};
 /// `compute_daily_runoff` still uses the legacy form. The total trim is
 /// 24 h under both conventions, so `T_days = T_hours/24 - 1` is unchanged.
 ///
+/// Why the last day is excluded: a window advanced by `tau` needs `tau`
+/// hours of routed output past the final store day, which were never
+/// routed (the q' feeding them is outside the window). Dropping exactly
+/// one day — front trim `tau` + back trim `24 - tau` — makes the SAME
+/// `T_days` days part of the training/testing set at every tau: tensor
+/// shapes, the obs pairing, and NSE(tau) sweeps all score an identical
+/// day sample regardless of the shift, and the trimmed length stays a
+/// multiple of 24 so pooling is an exact block mean. The alternative
+/// (score all days by routing one extra day of q' per window) is ~1% more
+/// training signal and needs +1-day window plumbing; deliberately not
+/// done — see the 2026-08-09 discussion in
+/// `docs/2026-08-06-tau-sweep-pilot-findings.md` §5i.
+///
 /// Returns `(G, T_days)` where `T_days = T_hours_trimmed // 24`.
 pub fn tau_trim_and_downsample<B: Backend>(
     predictions_hourly: Tensor<B, 2>,