Juniata single-catchment sample + deprecate ddr_match (corrected physics by default)

[taddyb]

Summary

Two tightly coupled changes:

1. examples/juniata/ — the ddrs mirror of DeepGroundwater/ddr#193. A self-contained quickstart on the Juniata River at Newport, PA (USGS 01567000, 8,657 km², 213 MERIT reaches), running on the byte-identical 8.9 MB bundle committed in the ddr repo: icechunk Q' + obs stores, attributes NetCDF, COO adjacency zarrs, one-row gage CSV, plus the attribute-statistics JSON (committed here because ddrs never recomputes statistics). Two commands, no HPC/S3/CUDA:

target/release/ddrs --config examples/juniata/ddrs.yaml plan
target/release/ddrs --config examples/juniata/ddrs.yaml run --workflow train-and-test --backend cpu

2. ddr_match deprecation. DDR ported ddrs's corrected physics (trapezoid-exact celerity β, Cunge-matched X, own-reach gauge readout) in DeepGroundwater/ddr#192 as a clean break, so the flag's default flips to false (corrected). Explicit ddr_match: true still parses with a deprecation WARN (needed to reproduce pre-#192 results and for use_cuda_graphs, whose captured kernel hardcodes the legacy 5/3 celerity). BREAKING: use_cuda_graphs: true without an explicit ddr_match: true now fails config load; config/merit_training.yaml flips to use_cuda_graphs: false.

Cross-implementation validation (same bundle, both directions)

test 1995-10-01 – 2010-09-30	NSE	KGE
ddrs routed (30 epochs, CPU, ~21 s)	0.790	0.881
DDR-Python routed (ddr#193 reference)	0.784	0.877
summed-Q' baseline (both implementations, to rounding)	0.695	0.819

The routed residual is only the window-sampling RNG stream (torch global RNG vs ChaCha12; ddrs 4-seed spread NSE 0.790–0.800). Invariant 1 re-anchored: the V1 sandbox fixture is regenerated from post-#192 DDR master and compare_ddr_sandbox reports ABSOLUTE MATCH on the corrected physics (max abs 1.5e-5 m³/s, max rel ~2e-7 — the f32 floor), which doubles as a bit-exactness proof of DDR's #192 port. The 2026-06-06 desktop-only-reference caveat is obsolete.

Test plan

• cargo test --no-fail-fast green under the new default (one legacy-physics fixture test now pins ddr_match: true; ddr_match_flag.rs asserts the new default; new juniata_bundle.rs contract tests never skip)
• compare_ddr_sandbox = ABSOLUTE MATCH against the regenerated post-#192 fixture
• Fresh-clone simulation (tracked files only, clean workspace): plan reproduces the baseline and run --workflow train-and-test --backend cpu reproduces NSE 0.790 / KGE 0.881 from the committed files alone

🤖 Generated with Claude Code