AORC2F/LSTM Q' campaign: experiment configs, findings, and eval corruption hardening

[taddyb]

Summary

This branch carries the AORC2F/LSTM Q' identifiability campaign: four new experiment configs, three findings docs (wave 1, cross-wave comparison, disagg-head sensitivity), comparison/plotting tooling, and the eval-corruption hardening that the campaign runs surfaced.

Experiment configs

• config/experiments/aorc2f_{distributed,lumped}_frozen_chunk1.yaml, lstm_daily_frozen_chunk1.yaml, lstm_hourly_native.yaml — the 4-source campaign templates.

Findings docs

• docs/2026-07-16-aorc2f-wave1-findings.md — wave 1 results.
• docs/2026-07-16-wave2-cross-wave-findings.md — cross-wave comparison, later amended to correct the backend-mismatch dismissal and test the double-routing hypothesis against the store (mixed result; controls still missing).
• docs/2026-07-16-disagg-head-sensitivity-findings.md — disaggregation-head sensitivity.
• docs/nh-qprime-store-contract.md updated.

Eval hardening (src/)

• Detect silently-corrupted eval chunks instead of writing bad predictions (src/training/eval.rs, src/bin/eval.rs), strengthened with a panic-hook flag so corruption can't slip through as a swallowed panic.
• Small supporting changes in src/cli/run.rs, src/cli/system.rs, src/cuda_graph/geometry_kernel.rs, src/sparse/cusparse.rs, src/data/error.rs.
• Doctest fix: fence pseudocode formulas so rustdoc doesn't compile them.

Tooling & skills

• scripts/generate_run_notebooks.py, scripts/hydrograph_comparison.py, scripts/wave_comparison_plots.py — per-run notebook generation and cross-wave comparison plots.
• .claude/skills/ddrs-run-and-operate/SKILL.md — documents the GPU eval OOM incident (cubecl worker-thread panic) and its fix.

Invariant check

Touches src/sparse/cusparse.rs and src/cuda_graph/ — per CLAUDE.md, cargo run --release --example compare_ddr_sandbox must still report ABSOLUTE MATCH before merge.

🤖 Generated with Claude Code