Restore global fiber workflows and native GPU/chunk tracing

.gitignore

@@ -1,5 +1,8 @@
 # Python venvs, IDE files, build artifacts
 /local/
+/task.md
+/experiments/
+/global-fiber-sgd
 **/.venv*/
 **/.idea/*
 **/__pycache__/

vesuvius/docs/examples/fiber_workflows/gt_width_regression.json

@@ -0,0 +1,18 @@
+{
+  "fiber_gt_convergence_probe": {
+    "source": "/data/fiber-layout/gt-initialization",
+    "reference": "/data/fiber-layout/gt-winding-reference.json",
+    "output_root": "/data/fiber-layout/results/gt-width-regression",
+    "devices": "0",
+    "gpu_manager": {
+      "exclusive": false,
+      "cpu_cores": 4,
+      "vram_bytes_per_gpu": 4294967296,
+      "label": "gt-regression"
+    },
+    "trials": [
+      {"name": "control", "checkpoint": "/data/fiber-layout/gt-initialization/solution.npz", "optimize": "joint", "steps": 100},
+      {"name": "perturbed", "checkpoint": "/data/fiber-layout/gt-initialization/solution.npz", "initial_width_scale": 1.2, "optimize": "joint", "steps": 100}
+    ]
+  }
+}

vesuvius/docs/examples/fiber_workflows/joint_gt.json

@@ -0,0 +1,10 @@
+{
+  "fiber_joint_gt_eval": {
+    "baseline_root": "/data/fiber-layout/prepared",
+    "base_config": "/configs/fiber-layout/base.json",
+    "gt_root": "/data/fiber-layout/gt",
+    "output_root": "/data/fiber-layout/results/joint",
+    "force_gt_enabled": true,
+    "devices": "1"
+  }
+}

vesuvius/docs/examples/fiber_workflows/json_pipeline.json

@@ -0,0 +1,13 @@
+{
+  "fiber_json_solve": {
+    "source_dir": "/data/fiber-layout/reviewed",
+    "afv_template": "/data/fiber-layout/metadata.afv",
+    "expected_manifest_tokens": ["YOUR_VOLUME", "YOUR_SCAN"],
+    "output_root": "/data/fiber-layout/prepared",
+    "stages": [
+      {"name": "extraction", "config_file": "/configs/fiber-layout/extraction.json"},
+      {"name": "orientation", "config_file": "/configs/fiber-layout/orientation.json"},
+      {"name": "layout", "config_file": "/configs/fiber-layout/layout.json"}
+    ]
+  }
+}

vesuvius/docs/examples/fiber_workflows/neighbors.json

@@ -0,0 +1,14 @@
+{
+  "fiber_neighbor_expansion": {
+    "seed_solve": "/data/fiber-layout/results/joint/solve",
+    "parent_tensors": "/data/fiber-layout/whole-sources/tensors.npz",
+    "afv": "/data/fiber-layout/sources.afv",
+    "gt_root": "/data/fiber-layout/gt",
+    "z_range": [30000, 50000],
+    "extract_template": "/configs/fiber-layout/extract.json",
+    "orient_template": "/configs/fiber-layout/orient.json",
+    "output_root": "/data/fiber-layout/expanded",
+    "confidence": 0.5,
+    "devices": "0"
+  }
+}

vesuvius/docs/examples/fiber_workflows/selected_dataset.json

@@ -0,0 +1,9 @@
+{
+  "fiber_selected_dataset_eval": {
+    "dataset": "/data/fiber-layout/selected",
+    "base_config": "/configs/fiber-layout/base.json",
+    "gt_root": "/data/fiber-layout/gt",
+    "output_root": "/data/fiber-layout/results/free",
+    "devices": "0"
+  }
+}

vesuvius/docs/examples/fiber_workflows/split_search.json

@@ -0,0 +1,21 @@
+{
+  "fiber_split_weight_search": {
+    "base_config": "/configs/fiber-layout/base.json",
+    "dataset": "/data/fiber-layout/selected",
+    "baseline_root": "/data/fiber-layout/prepared",
+    "gt_root": "/data/fiber-layout/gt",
+    "output_root": "/data/fiber-layout/search",
+    "devices": ["0", "1"],
+    "initial_parameters": {"rejection_cost": 1, "continuation_weight": 1, "split_min_perpendicular": 3},
+    "weights": ["rejection_cost", "continuation_weight", "split_min_perpendicular"],
+    "bounds": {"rejection_cost": [0, 10], "continuation_weight": [0, 10], "split_min_perpendicular": [2, 6]},
+    "factor_grids": [[0.5, 2], [0.9, 1.1]],
+    "discrete_values": {"split_min_perpendicular": [2, 3, 6]},
+    "proposal_groups": [{
+      "name": "joint_costs",
+      "axes": [{"parameters": ["rejection_cost", "continuation_weight"], "mode": "factor", "values": [0.5, 2]}]
+    }],
+    "accepted_improvements": null,
+    "background": false
+  }
+}

vesuvius/docs/fiber_layout.md

@@ -2,8 +2,10 @@
 
 This extracts the existing AFV-to-constraint and fiber-layout implementations.
 Inputs and earlier stage solutions are immutable; preparation records source
-identities, hashes and coordinate conventions. Experiment matrices, run-specific
-configs/data and layout-to-Lasagna surface reconstruction are not included.
+identities, hashes and coordinate conventions. Preparation, search, experiment
+matrices, diagnostics and tifxyz surface export are documented in
+[fiber-layout workflows](fiber_workflows.md). Run-specific configs and data are
+external; portable usage samples are included.
 
 ## Native fixed evidence
 
@@ -53,6 +55,7 @@ see [the N/U coordinate contract](fiber_layout_coordinates.md). Integer-label
 curves are not evaluated at a fractional distribution center, and old saved
 geometry is not silently reinterpreted. Reference alignment is global, not a
 per-fiber fit. Ordered evidence and family conventions are persisted.
+An [initialization-width prior must not be reintroduced](fiber_layout_no_width_prior.md).
 
 
 `fiber_gt_constraint_score.run(settings)` can build cross-reference event caches

vesuvius/docs/fiber_layout_benchmark.md

@@ -131,21 +131,8 @@ Tests mock GPU launches. Actual data replay requires the prepared tensors,
 initializers/samples, frozen GT, cached cross-events, and historical reports;
 these large artifacts are not bundled. No installation or bootstrap is done.
 
-Real-data compatibility check (2026-10-22): replaying the historical accepted
-candidate `a838956dd140d329e78f` through the new scorer reproduced all three
-recorded perpendicular/parallel scores and sign counts (L1/correctness tolerance
-1e-10). Summed perpendicular signs: 38; selectable GT retained length: 95.315808%;
-rank score: 39.867469; summed perpendicular L1: 1.023098. This checks the scoring
-API against real checkpoints, **not** convergence of new solves on main.
-
-Review-fix validation (2026-10-22): all 14 recorded final loss terms and all
-other final scientific fields match for each of the three historical checkpoints.
-The same scores above are reproduced with the stricter checks. Auditing the
-170/171/170 source settings and the solver's prepared configs found no scientific
-setting changes versus the historical executed configs; differences are output
-paths, GPU lanes/labels, and code-provenance metadata. The L-BFGS, geometry and
-loss implementations were also compared against the historical solver revision
-`4a35bf9e18c3` (and the last experiment branch `d8f8c3595`): objective calculations
-are unchanged. Changes in the merged entrypoint modules concern launch dispatch,
-moved imports and removal of experiment-only dispatch/audits. No fresh GPU solve
-was run for this review-fix check.
+For a read-only compatibility check against caller-supplied saved checkpoints,
+set `FIBER_LAYOUT_REFERENCE_SPEC` to a benchmark specification and run
+`vesuvius/tests/test_fiber_workflows_real_data.py`. This recomputes scores and
+full-objective terms without running optimization. Keep experiment-specific
+measurements and checkpoint identities with the external results, not here.

vesuvius/docs/fiber_layout_no_width_prior.md

@@ -0,0 +1,14 @@
+# No initialization-width prior
+
+Initialization widths are guesses, not observations. **Do not re-add a loss
+pulling fitted winding widths toward initialization**, including a radius-based
+initialization. Do not expose such a term to tuning or restore it from a recipe.
+
+The former `width_prior * mean((width / initial_width - 1)^2)` is deleted from
+the soft/probability objective, legacy layout solver and diagnostic objective.
+Nonzero legacy settings are rejected. Zero-valued historical settings remain
+readable but have no loss implementation.
+
+Width smoothness and the soft minimum-width penalty are separate, valid terms.
+Initialization may supply starting values and numerical scales, never a target
+fitted width. Historic results with the deleted prior are not valid baselines.