daily.yml: pre-pull per-year AASHTO crashes.parquet deps

Runs 28915961775 + 28916371434 (both 2026-07-08) failed at the
"Re-match NJSP↔NJDOT" step: DVX saw `njdot/data/{2023,2024,2025}/
crashes.parquet` missing on disk, decided to re-run their commit
stages, and fell through to `njdot/aashto/normalize.py: not found`
(a latent shebang / `uv`-on-PATH issue in the re-run path). Cache is
now populated on S3 (pushed manually), but DVX doesn't attempt to
fetch when the output file is absent from the workspace — it just
runs the cmd.

Same fix pattern as `persons.parquet` / `vehicles.parquet` above:
pull the per-year outputs before `dvx run` so `is_output_fresh`
returns true and the re-execute path isn't taken. Root-cause fix for
the shebang path can come separately.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: ryan-williams

.github/workflows/daily.yml

@@ -79,6 +79,16 @@ jobs:
           dvx pull njdot/data/2023/vehicles.parquet.dvc
           dvx pull njdot/data/2024/vehicles.parquet.dvc
           dvx pull njdot/data/2025/vehicles.parquet.dvc
+          # Per-year AASHTO `crashes.parquet` — same rationale as persons /
+          # vehicles above. Consumed transitively by
+          # `aashto_combined_crashes.parquet` (which is pulled below). Without
+          # these on disk, DVX sees the intermediate stage's output-fresh
+          # check pass but the *upstream* per-year stage's output missing,
+          # tries to re-run `normalize.py`, and hits the shebang-not-found
+          # path (see #127-adjacent notes).
+          dvx pull njdot/data/2023/crashes.parquet.dvc
+          dvx pull njdot/data/2024/crashes.parquet.dvc
+          dvx pull njdot/data/2025/crashes.parquet.dvc
           dvx pull njsp/data/crash-log.parquet.dvc
           # NJSP outputs of `njsp update_pqts` (consumed by harmonize, csvs,
           # projections, match, three-way-match, d1-import).