Add calibration tolerance experiment app

Author: anth-volk

modal_app/README.md

@@ -27,6 +27,8 @@ modal run modal_app/remote_calibration_runner.py --branch <branch> --epochs <n>
 | `--push-results` | `False` | Upload weights, blocks, and logs to HuggingFace |
 | `--trigger-publish` | `False` | Fire `repository_dispatch` to trigger the Publish workflow |
 | `--target-config` | (none) | Target configuration name |
+| `--target-policy` | (none) | Target tolerance policy YAML path |
+| `--target-tolerance-mode` | (none) | Final tolerance behavior: `enforce` or `evaluate` |
 | `--beta` | (none) | L0 relaxation parameter |
 | `--lambda-l0` | (none) | L0 penalty weight |
 | `--lambda-l2` | (none) | L2 penalty weight |
@@ -78,6 +80,8 @@ Every run produces these local files (whichever the calibration script emits):
 - **unified_diagnostics.csv** — Final per-target diagnostics
 - **calibration_log.csv** — Per-target metrics across epochs (requires `--log-freq`)
 - **unified_run_config.json** — Run configuration and summary stats
+- **calibration_target_policy.jsonl** — Resolved per-target tolerance policy
+- **calibration_target_policy_summary.json** — Target tolerance summary
 
 ## Artifact Upload to HuggingFace
 
@@ -90,10 +94,36 @@ atomic commit after writing them locally:
 | `calibration_log.csv` | `calibration/logs/calibration_log.csv` |
 | `unified_diagnostics.csv` | `calibration/logs/unified_diagnostics.csv` |
 | `unified_run_config.json` | `calibration/logs/unified_run_config.json` |
+| `calibration_target_policy.jsonl` | `calibration/logs/calibration_target_policy.jsonl` |
+| `calibration_target_policy_summary.json` | `calibration/logs/calibration_target_policy_summary.json` |
 
 Each upload overwrites the previous files. HF git history provides implicit
 versioning — browse past commits to see earlier runs.
 
+## Tolerance Experiment App
+
+`modal_app/calibration_tolerance_experiment.py` runs an isolated full regional
+calibration experiment from the current Modal volume inputs. It copies
+`policy_data.db` and `source_imputed_stratified_extended_cps.h5` into
+`/pipeline/epsilon_calibration_runs/{run_id}`, builds a run-scoped calibration
+package, fits weights with `--target-tolerance-mode evaluate`, and writes
+diagnostics, resolved target-policy artifacts, and `experiment_summary.json`.
+It does not upload Hugging Face artifacts or trigger H5 publication.
+
+Deploy the app when you want to launch repeat experiments from Modal rather
+than from a local CLI session:
+
+```bash
+modal deploy modal_app/calibration_tolerance_experiment.py
+```
+
+```bash
+modal run modal_app/calibration_tolerance_experiment.py::main \
+  --branch codex/epsilon-insensitive-calibration \
+  --run-id epsilon-smoke \
+  --epochs 1000
+```
+
 ## Triggering the Publish Workflow
 
 The `--trigger-publish` flag fires a `repository_dispatch` event