fix conv kernel support truncation...

.claude/skills/benchmark/benchmark_gir.py

@@ -74,39 +74,79 @@ def _extract_float(src, key):
 
 
 #
-def _parse_dynamics_from_notebook(notebook_path):
-    """Extract add_time_dependence kwargs from example.ipynb.
+def _iter_call_args(src, func_name):
+    """Yield the argument text of each ``func_name(...)`` call in *src*.
+
+    Matches parentheses so multiple calls per cell stay separate (a
+    cell-level regex would mix kwargs of neighboring calls).
+    """
+
+    start = 0
+    while True:
+        idx = src.find(func_name + "(", start)
+        if idx == -1:
+            return
+        depth = 0
+        for j in range(idx + len(func_name), len(src)):
+            if src[j] == "(":
+                depth += 1
+            elif src[j] == ")":
+                depth -= 1
+                if depth == 0:
+                    yield src[idx + len(func_name) + 1 : j]
+                    start = j + 1
+                    break
+        else:
+            return
+
+
+#
+def _parse_calls_from_notebook(notebook_path):
+    """Extract add_par_profile / add_time_dependence kwargs from example.ipynb.
 
     Returns
     -------
-    list of dict
-        Each dict has keys: target_model, target_parameter,
+    profile_calls : list of dict
+        add_par_profile kwargs: target_model, target_parameter,
+        profile_yaml, profile_model.
+    dynamics_calls : list of dict
+        add_time_dependence kwargs: target_model, target_parameter,
         dynamics_yaml, dynamics_model, and optionally frequency.
     """
 
     with notebook_path.open() as f:
         nb = json.load(f)
 
-    calls = []
+    profile_calls = []
+    dynamics_calls = []
     for cell in nb["cells"]:
         if cell["cell_type"] != "code":
             continue
         src = "".join(cell["source"])
-        if "add_time_dependence" not in src:
-            continue
 
-        call = {
-            "target_model": _extract_str(src, "target_model"),
-            "target_parameter": _extract_str(src, "target_parameter"),
-            "dynamics_yaml": _extract_str(src, "dynamics_yaml"),
-            "dynamics_model": _extract_str_or_list(src, "dynamics_model"),
-        }
-        freq = _extract_float(src, "frequency")
-        if freq is not None:
-            call["frequency"] = freq
-        calls.append(call)
+        for args in _iter_call_args(src, "add_par_profile"):
+            profile_calls.append(
+                {
+                    "target_model": _extract_str(args, "target_model"),
+                    "target_parameter": _extract_str(args, "target_parameter"),
+                    "profile_yaml": _extract_str(args, "profile_yaml"),
+                    "profile_model": _extract_str_or_list(args, "profile_model"),
+                }
+            )
+
+        for args in _iter_call_args(src, "add_time_dependence"):
+            call = {
+                "target_model": _extract_str(args, "target_model"),
+                "target_parameter": _extract_str(args, "target_parameter"),
+                "dynamics_yaml": _extract_str(args, "dynamics_yaml"),
+                "dynamics_model": _extract_str_or_list(args, "dynamics_model"),
+            }
+            freq = _extract_float(args, "frequency")
+            if freq is not None:
+                call["frequency"] = freq
+            dynamics_calls.append(call)
 
-    return calls
+    return profile_calls, dynamics_calls
 
 
 # ------------------------------------------------------------------
@@ -140,7 +180,10 @@ def load_example(example_num, *, add_dynamics=True):
     -------
     file : File
     dynamics_calls : list of dict
-        Parsed add_time_dependence kwargs from the notebook.
+        Parsed add_time_dependence kwargs from the notebook ("2D" model
+        only).
+    profile_calls : list of dict
+        Parsed add_par_profile kwargs already attached to the model.
     """
 
     folder = _find_example_folder(example_num)
@@ -152,23 +195,35 @@ def load_example(example_num, *, add_dynamics=True):
     energy = np.loadtxt(data_dir / "energy.csv")
     time_ax = np.loadtxt(data_dir / "time.csv")
     data = np.loadtxt(data_dir / "data.csv", delimiter=",")
+    aux_path = data_dir / "aux_axis.csv"
+    aux_axis = np.loadtxt(aux_path) if aux_path.exists() else None
 
     file = File(
         parent_project=project,
         name="bench",
         data=data,
         energy=energy,
         time=time_ax,
+        aux_axis=aux_axis,
     )
     file.load_model(model_yaml="models_energy.yaml", model_info="2D")
 
-    dynamics_calls = _parse_dynamics_from_notebook(folder / "example.ipynb")
+    profile_calls, dynamics_calls = _parse_calls_from_notebook(folder / "example.ipynb")
+    # The notebooks also attach to their baseline/SbS models; only calls
+    # targeting the benchmarked "2D" model apply here.
+    profile_calls = [c for c in profile_calls if c["target_model"] == "2D"]
+    dynamics_calls = [c for c in dynamics_calls if c["target_model"] == "2D"]
+
+    # Profiles are model structure, not time dependence: attach always,
+    # and before dynamics (dynamics may target a profile parameter).
+    for call in profile_calls:
+        file.add_par_profile(**call)
 
     if add_dynamics:
         for call in dynamics_calls:
             file.add_time_dependence(**call)
 
-    return file, dynamics_calls
+    return file, dynamics_calls, profile_calls
 
 
 #
@@ -333,7 +388,7 @@ def _snapshot(label):
         print(f"  {label:32s}{call_count[0]:6d}  (+{call_count[0] - prev})")
 
     try:
-        file, dynamics_calls = load_example(example_num, add_dynamics=False)
+        file, dynamics_calls, _ = load_example(example_num, add_dynamics=False)
         file.define_baseline(
             time_start=0, time_stop=10, time_type="ind", show_plot=False
         )
@@ -397,7 +452,7 @@ def timed_sched(*args, **kwargs):
     graph_ir.schedule_2d = timed_sched
 
     try:
-        file, dynamics_calls = load_example(example_num, add_dynamics=False)
+        file, dynamics_calls, _ = load_example(example_num, add_dynamics=False)
         file.define_baseline(
             time_start=0, time_stop=10, time_type="ind", show_plot=False
         )
@@ -454,7 +509,7 @@ def capture_par_variability(example_num, *, n_starts=4):
     redchis = []
 
     for run in range(n_starts + 1):
-        file, dynamics_calls = load_example(example_num, add_dynamics=False)
+        file, dynamics_calls, _ = load_example(example_num, add_dynamics=False)
         file.define_baseline(
             time_start=0, time_stop=10, time_type="ind", show_plot=False
         )
@@ -563,7 +618,7 @@ def bench_fit(example_num, dynamics_calls, *, n_reps=3):
             ("fit_model_gir", gir_times),
             ("fit_model_mcp", mcp_times),
         ]:
-            file, _ = load_example(example_num, add_dynamics=False)
+            file, _, _ = load_example(example_num, add_dynamics=False)
             file.define_baseline(
                 time_start=0, time_stop=10, time_type="ind", show_plot=False
             )
@@ -682,11 +737,15 @@ def bench_fit(example_num, dynamics_calls, *, n_reps=3):
         folder = _find_example_folder(args.example)
         print(f"Example: {folder.name}")
 
-        file, dynamics_calls = load_example(args.example)
+        file, dynamics_calls, profile_attaches = load_example(args.example)
         model = file.model_active
         assert model is not None
         graph = build_graph(model)
         print(f"  lowerable: {can_lower_2d(graph)}")
+        for call in profile_attaches:
+            target_parameter = call["target_parameter"]
+            profile_model = call["profile_model"]
+            print(f"  profile:   {target_parameter} <- {profile_model}")
         for call in dynamics_calls:
             target_parameter = call["target_parameter"]
             dynamics_model = call["dynamics_model"]

CHANGELOG.md

@@ -7,6 +7,29 @@ and this project adheres to [Semantic Versioning](https://semver.org/).
 This file is maintained using the shared changelog workflow in
 [`docs/ai/changelog.md`](docs/ai/changelog.md).
 
+## [0.10.2] - 2026-07-10
+
+### Added
+
+- **Diverging colormap for residual maps**: new `z_colormap_res` plot setting (default `'RdBu_r'`). 2D data+fit+residual panels now render the residual with its own zero-centered diverging colormap and symmetric color limits, instead of reusing the data colormap.
+- **New plot styling settings**: `refline_color` / `refline_style` for `vlines`/`hlines` reference lines (unified default: grey dotted — previously inconsistent between 1D and 2D plots) and `panel_size` for multi-panel grid plots; `plot_2d_grid` gained a `columns` argument.
+- **Every `PlotConfig` field is now settable via `project.yaml`**, with tuple-valued fields (`x_lim`, `panel_size`, ...) accepting YAML lists; a coverage test guards future fields.
+- `results_to_fit_2d` accepts `parameter_names` to select and order DataFrame columns, so extra non-parameter columns (e.g. from `results_to_df` output) cannot silently corrupt the reconstructed 2D fit.
+
+### Changed
+
+- **Silent mode is honored throughout**: with `show_output=0`, MCMC no longer prints its banner/progress bar or leaves walker/corner figures open, and `fit_2d`/`fit_slice_by_slice` no longer display timing and parameter tables. A broken `project.yaml` now raises `ValueError` instead of silently falling back to defaults.
+- **Clear errors instead of silent misbehavior**: fits reject NaN/Inf inside the fit window with a message naming the count and pointing to `set_fit_limits()`; `describe`/`define_baseline`/`set_fit_limits` raise when axes are missing instead of fabricating index axes; convolution on a single-point time axis, zero-sum convolution kernels, unknown `Simulator` noise types (constructor and setter), unknown sweep distribution types, empty parameter sweeps, and dynamics-model expressions referencing parameters outside their own dynamics model all raise immediately with actionable messages; a `t_vary` parameter without a dynamics model raises instead of returning `-1.0`; `LinBack` ordering errors suggest setting bounds on `xStart`/`xStop`.
+- **Performance**: the compiled 2D evaluator avoids per-instruction array allocation and copies (preallocated profile buffers, hoisted profiled-op parameter sources), `my_conv` no longer builds a discarded padded axis, MCMC walker/corner figures are only constructed when shown or saved, and parameter sweeps write to one open HDF5 file instead of reopening it per configuration.
+
+### Fixed
+
+- **Convolution kernel support no longer truncates during fits**: the IRF kernel time axis was built once from the initial width parameter and never rebuilt, silently truncating the kernel and biasing the fitted width once it grew past its initial value. Both evaluation paths now rebuild the kernel support from current parameter values on every evaluation.
+- **Partial-range 2D evaluation used wrong time indices**: `Model.create_value_2d(t_ind=[start, stop])` computed dynamics for `t[0:stop-start]` instead of `t[start:stop]`.
+- `Simulator.plot_comparison` on a fresh simulator now auto-simulates with current settings instead of failing.
+- `sign_change` no longer hangs on all-zero input.
+- `plot_1d` with `y_norm=1` no longer produces all-NaN plots for constant traces.
+
 ## [0.10.0] - 2026-07-08
 
 ### Added

CLAUDE.md

@@ -1,6 +1,7 @@
 # General behavior
 
 - **Confidence Rule:** Do not make changes until you have 95% confidence. Understand the relevant files before editing and ask follow-up questions until you reach this threshold or when tradeoffs are non-obvious.
+- **Commits:** Never commit or rewrite history unless explicitly asked. Always show the exact commit message and wait for approval before committing. Never add AI attribution trailers (`Co-Authored-By` etc.) to commit messages.
 - **Context Discipline:** Monitor context usage. At 60% usage (or if it starts getting tight), summarize progress and prompt me to `/compact` or `/clear`.
 - **Token Efficiency:** Be concise. Reference file paths and line numbers rather than quoting large code blocks.
 - **Subagent Protocol:** Use subagents for repo-wide scans, parallel research, or scanning large directories. Instruct them to return only concise summaries to keep the main context window lean.
@@ -41,7 +42,7 @@
 - **Pattern:** Use plain pytest. Avoid `unittest.TestCase` and fixtures; prefer explicit helper builders named by intent.
 - **API Usage:** Use the public API (`Project`, `File.load_model`, etc.) in tests to avoid masking bugs by skipping validation or setup. Use internals only for pure-math unit tests or explicit invariant checks.
 - **Execution:** Run `pytest -q`. Keep YAML test assets in `tests/models/`.
-- **Plots:** Always suppress plot display in tests: pass `show_plot=False` where available, or `save_img=-2`.
+- **Plots:** Always suppress plot display in tests: pass `show_plot=False` where available, or `save_img=-2`. Exception: figure-inspection tests that assert on the live axes cannot pass `save_img=-2` (it closes the figure); they rely on the module-level Agg backend and must call `plt.close("all")` after the assertions.
 - **Type Guards:** When `assert x is not None` narrows an `X | None` type, add a `# type guard` comment.
 - **Variable Naming:** For variables derived from registry parameters or components, keep original casing (e.g., `SD = 2.0`, `c_Shirley = Component("Shirley")`). Name derived variables as `{par}_{qualifier}` (e.g., `A_early`, `mean_A`).