Release v0.3.0-alpha.19: resolve var_ids before pulling samples

Fixes bug #23 discovered while end-to-end testing alpha.18 on a real
workbook. After a fresh run_simulation, get_sensitivity_ranking
returned empty on the first call and full ranking on the immediate
retry. Diagnostic: output lookup succeeded, output samples loaded,
then every input lookup against the same handle returned None.

Contract finding about MRService.dll: MRLIB_GetModelData appears to
leave the handle in a state where subsequent MRLIB_GetModelVarID
calls return None. The call sequence within one open handle must be
all GetModelVarID calls first, then all GetModelData calls.
Interleaving is unsafe.

Fix: every reader that interleaves lookups with sample fetches now
resolves all var_ids first, then pulls samples. Touches:
- ResultsReader.get_sensitivity_ranking
- ResultsReader.get_simulation_results
- ResultsReader.get_correlation_matrix

get_samples (single name, single fetch) is fine as-is.

Verified live: first sensitivity-ranking call on NPV_of_a_capital
_investment complete.xlsx now returns 6 driver entries (top driver
Market growth, r = +0.72) instead of an empty list.

399 unit tests still pass — the regression test is the integration
smoke run since the bug only manifests against the real DLL.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: vosesoft

CHANGELOG.md

@@ -4,6 +4,28 @@ All notable changes to ModelRisk MCP. Follows [Keep a Changelog](https://keepach
 
 ## [Unreleased]
 
+## [0.3.0-alpha.19] — 2026-05-22
+
+Fixes the bug-#23 lookup-after-samples regression discovered while end-to-end testing alpha.18 against a real workbook: `get_sensitivity_ranking` returned empty on the first call after `run_simulation`, then worked on the second identical call. The diagnostic trace was unambiguous — the output looked up fine, its samples loaded, then every input lookup against the same handle returned None.
+
+### Fixed
+
+- **MRLIB_GetModelData poisons subsequent MRLIB_GetModelVarID calls on the same handle.** Resolution: every reader that interleaves name lookups with sample fetches now resolves ALL var_ids first, THEN pulls samples. Applies to:
+  - `ResultsReader.get_sensitivity_ranking` — was failing on the first call after a fresh simulation (output looked up, output samples loaded, all inputs then refused to resolve). Now: output lookup → all input lookups → all sample fetches → ranking.
+  - `ResultsReader.get_simulation_results` — same risk on multi-output calls. Same fix.
+  - `ResultsReader.get_correlation_matrix` — same risk on multi-name correlation requests. Same fix.
+  - `ResultsReader.get_samples` (single name, single fetch — no change needed).
+
+This is a *contract* finding about MRService.dll: the call sequence within one open handle must be all `GetModelVarID` calls first, then all `GetModelData` calls. Inverting them or interleaving is unsafe. Worth flagging upstream to the ModelRisk SDK team — and worth knowing for any future readers that touch the same surface.
+
+### Why this matters end-to-end
+
+Without alpha.19, the user's first sensitivity-ranking call after a sim returned silently empty. The LLM would tell them "no drivers detected" — completely wrong on a model that clearly has Spearman correlations up to +0.72. After alpha.19 the first call works correctly. Verified live against the `NPV_of_a_capital_investment complete.xlsx` workbook: 6 driver entries returned, top driver Market growth (r = +0.72), bottom three in noise territory.
+
+### Tests
+
+399 unit tests still pass — the bug only manifests against the real DLL, so the regression test is the integration smoke run.
+
 ## [0.3.0-alpha.18] — 2026-05-22
 
 Targeted experiment for the empty-`.vmrs` blocker surfaced by alpha.17's post-condition verification. The bridge correctly detected that `VoseStartSimulCustom12 + VoseGetDataSZ12` was producing `.vmrs` files with zero registered outputs — sim ran, file existed, but no variable metadata. Ribbon-driven simulations on the same workbook worked fine, suggesting the ribbon path threads an option the headless XLL path skipped.

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "modelrisk-mcp"
-version = "0.3.0a18"
+version = "0.3.0a19"
 description = "Open MCP server bridging Anthropic Claude (and any MCP-compatible client) with the ModelRisk Excel add-in."
 readme = "README.md"
 requires-python = ">=3.11"

server.json

@@ -3,12 +3,12 @@
   "name": "io.github.vosesoftware/modelrisk-mcp",
   "title": "ModelRisk",
   "description": "Read, build, fit, and run Monte Carlo risk models in Excel through Vose Software's ModelRisk.",
-  "version": "0.3.0a18",
+  "version": "0.3.0a19",
   "packages": [
     {
       "registryType": "pypi",
       "identifier": "modelrisk-mcp",
-      "version": "0.3.0a18",
+      "version": "0.3.0a19",
       "transport": {
         "type": "stdio"
       }

src/modelrisk_mcp/__init__.py

@@ -1 +1 @@
-__version__ = "0.3.0a18"
+__version__ = "0.3.0a19"

src/modelrisk_mcp/bridge/results.py

@@ -124,10 +124,18 @@ def get_simulation_results(
                 # this fallback just won't return anything sensible here.
                 # Use the read_vmrs tool with explicit output_names for now.
                 return []
+            # Bug #23 (alpha.19): MRLIB_GetModelData appears to put the
+            # handle into a state where subsequent MRLIB_GetModelVarID
+            # calls return None — so we resolve EVERY name to a var_id
+            # before pulling any samples. Without this, a fresh-from-
+            # simulation .vmrs would silently lose all but the first
+            # output (lookup → samples → lookup → fails → continue → ...).
+            resolved: list[tuple[str, int]] = []
             for name in wanted:
                 var_id = self._lookup_var_id(handle, name)
-                if var_id is None:
-                    continue
+                if var_id is not None:
+                    resolved.append((name, var_id))
+            for name, var_id in resolved:
                 samples = handle.get_samples(var_id)
                 if not samples:
                     continue
@@ -159,12 +167,18 @@ def get_correlation_matrix(
             return CorrelationMatrix()
         name_list = list(names)
         with self._mrservice.open_vmrs(vmrs) as handle:
-            arrays: list[np.ndarray] = []
-            ordered: list[str] = []
+            # Bug #23 (alpha.19): resolve every name to a var_id BEFORE
+            # touching get_samples. See get_simulation_results for the
+            # full backstory — MRLIB_GetModelData poisons subsequent
+            # MRLIB_GetModelVarID lookups on the same handle.
+            resolved: list[tuple[str, int]] = []
             for name in name_list:
                 var_id = self._lookup_var_id(handle, name)
-                if var_id is None:
-                    continue
+                if var_id is not None:
+                    resolved.append((name, var_id))
+            arrays: list[np.ndarray] = []
+            ordered: list[str] = []
+            for name, var_id in resolved:
                 samples = handle.get_samples(var_id)
                 if not samples:
                     continue
@@ -250,17 +264,28 @@ def get_sensitivity_ranking(
         if vmrs is None:
             raise SimulationFailedError("No .vmrs available.")
         with self._mrservice.open_vmrs(vmrs) as handle:
+            # Bug #23 (alpha.19): resolve EVERY name to a var_id before
+            # touching get_samples. MRLIB_GetModelData appears to leave
+            # the handle in a state where subsequent
+            # MRLIB_GetModelVarID calls return None — first observed
+            # on a fresh-from-simulation .vmrs where the output looked
+            # up fine, its samples loaded, then every input lookup
+            # returned None and the sensitivity ranking came back
+            # empty. Resolving all IDs first sidesteps the issue.
             out_id = self._lookup_var_id(handle, output_name)
             if out_id is None:
                 raise SimulationFailedError(
                     f"Output {output_name!r} not found in .vmrs."
                 )
-            out_samples = np.asarray(handle.get_samples(out_id), dtype=float)
-            entries: list[SensitivityEntry] = []
+            resolved_inputs: list[tuple[str, int]] = []
             for in_name in input_names:
                 in_id = self._lookup_var_id(handle, in_name)
-                if in_id is None:
-                    continue
+                if in_id is not None:
+                    resolved_inputs.append((in_name, in_id))
+
+            out_samples = np.asarray(handle.get_samples(out_id), dtype=float)
+            entries: list[SensitivityEntry] = []
+            for in_name, in_id in resolved_inputs:
                 in_samples = np.asarray(handle.get_samples(in_id), dtype=float)
                 n = min(in_samples.size, out_samples.size)
                 if n < 2:

tests/unit/test_server_boot.py

@@ -5,7 +5,7 @@
 
 
 def test_version_is_set() -> None:
-    assert __version__ == "0.3.0a18"
+    assert __version__ == "0.3.0a19"
 
 
 def test_server_name() -> None: