[UsatovPavel]

Jira: https://goal-jira.atlassian.net/browse/HACK-145

Adds the 4-class incident pipeline (run v4) in retraining/notebooks/incident4/. Existing notebooks, app.py, inference/, data/ and CI are not touched.

What's inside

• Notebooks 1–3 from the v4 run, with outputs and execution counts stripped:
  • target by справочник_состояний («вариант 1») and the «Не замкнут» fix;
  • intra-second look-ahead fix and dedup of journal rows;
  • 80/10/10 split by channel within each class, plus a June 2026 blind set;
  • June censoring, and GAS 0.15 / INTRUSION 0.5 thinning in train only;
  • feature selection by val macro PR-AUC.
• train_v4.py trains in segments of 100 trees and stops early on val macro PR-AUC.
• eval_v4.py runs the one-shot blind evaluation against v3, prod and the baseline rules.
• check_lookahead.py brute-force checks the look-ahead fix.
• pipeline_v4.sh runs everything in order and best_params.json holds the v1 params.
• README.md (Russian) covers the differences from the original notebooks, how to run, resources and the blind-test results with caveats.

VM-specific paths are now parameters:

• the run dir is RUN_DIR (argument or env), defaulting to the git-ignored retraining/artifacts/runs/incident4_v4;
• comparison models and chunk maps come from COMPARE_DIR / V3_MODEL / PROD_MODEL / V3_CHUNK_MAP / V2_CHUNK_MAP;
• the private notebook runner is replaced by jupyter nbconvert --execute.

Model logic is unchanged. Model binaries are not committed.

Size

About 1880 added lines, mostly notebook JSON. This is above the usual 1000/1500 limit, which was agreed with the owner for this single PR.

Checks

• Notebooks validated with nbformat and have no outputs.
• py_compile passes on the scripts and bash -n passes on the shell script.
• Grep finds no user paths, hosts or secrets.
• The repo tests were not run locally (pytest is not in the venv). The PR only adds files that nothing imports, and CI runs the suite.

🤖 Generated with Claude Code