diff --git a/.config/nextest.toml b/.config/nextest.toml
index e3a8de680..01d7536c4 100644
--- a/.config/nextest.toml
+++ b/.config/nextest.toml
@@ -1,13 +1,3 @@
 [profile.default]
 # Mark tests that take longer than 10s as slow
 slow-timeout = "10s"
-
-[[profile.default.overrides]]
-# Some of these tests have a non-determinstic snapshot
-filter = 'binary(pip_install_scenarios)'
-retries = 2
-
-[[profile.no-retry.overrides]]
-# An optional profile to avoid retries
-filter = 'binary(pip_install_scenarios)'
-retries = 0
diff --git a/crates/puffin-resolver/src/error.rs b/crates/puffin-resolver/src/error.rs
index c84a3181d..32b17c6fa 100644
--- a/crates/puffin-resolver/src/error.rs
+++ b/crates/puffin-resolver/src/error.rs
@@ -2,6 +2,7 @@ use std::collections::BTreeSet;
 use std::convert::Infallible;
 use std::fmt::Formatter;
 
+use dashmap::DashSet;
 use indexmap::IndexMap;
 use pubgrub::range::Range;
 use pubgrub::report::{DefaultStringReporter, DerivationTree, Reporter};
@@ -167,6 +168,7 @@ impl NoSolutionError {
     pub(crate) fn with_available_versions(
         mut self,
         python_requirement: &PythonRequirement,
+        visited: &DashSet<PackageName>,
         package_versions: &OnceMap<PackageName, VersionMap>,
     ) -> Self {
         let mut available_versions = IndexMap::default();
@@ -186,15 +188,21 @@ impl NoSolutionError {
                     );
                 }
                 PubGrubPackage::Package(name, ..) => {
-                    if let Some(entry) = package_versions.get(name) {
-                        let version_map = entry.value();
-                        available_versions.insert(
-                            package.clone(),
-                            version_map
-                                .iter()
-                                .map(|(version, _)| version.clone())
-                                .collect(),
-                        );
+                    // Avoid including available versions for packages that exist in the derivation
+                    // tree, but were never visited during resolution. We _may_ have metadata for
+                    // these packages, but it's non-deterministic, and omitting them ensures that
+                    // we represent the state of the resolver at the time of failure.
+                    if visited.contains(name) {
+                        if let Some(entry) = package_versions.get(name) {
+                            let version_map = entry.value();
+                            available_versions.insert(
+                                package.clone(),
+                                version_map
+                                    .iter()
+                                    .map(|(version, _)| version.clone())
+                                    .collect(),
+                            );
+                        }
                     }
                 }
             }
diff --git a/crates/puffin-resolver/src/resolver/mod.rs b/crates/puffin-resolver/src/resolver/mod.rs
index 435c4a796..998600398 100644
--- a/crates/puffin-resolver/src/resolver/mod.rs
+++ b/crates/puffin-resolver/src/resolver/mod.rs
@@ -3,7 +3,7 @@
 use std::sync::Arc;
 
 use anyhow::Result;
-use dashmap::DashMap;
+use dashmap::{DashMap, DashSet};
 use futures::channel::mpsc::UnboundedReceiver;
 use futures::{pin_mut, FutureExt, StreamExt};
 use itertools::Itertools;
@@ -68,6 +68,8 @@ pub struct Resolver<'a, Provider: ResolverProvider> {
     index: &'a InMemoryIndex,
     /// A map from [`PackageId`] to the `Requires-Python` version specifiers for that package.
     incompatibilities: DashMap<PackageId, VersionSpecifiers>,
+    /// The set of all registry-based packages visited during resolution.
+    visited: DashSet<PackageName>,
     editables: FxHashMap<PackageName, (LocalEditable, Metadata21)>,
     reporter: Option<Arc<dyn Reporter>>,
     provider: Provider,
@@ -167,6 +169,7 @@ impl<'a, Provider: ResolverProvider> Resolver<'a, Provider> {
         Self {
             index,
             incompatibilities: DashMap::default(),
+            visited: DashSet::default(),
             selector,
             allowed_urls,
             project: manifest.project,
@@ -223,7 +226,7 @@ impl<'a, Provider: ResolverProvider> Resolver<'a, Provider> {
                     if let ResolveError::NoSolution(err) = err {
                         ResolveError::NoSolution(
                             err
-                            .with_available_versions(&self.python_requirement, &self.index.packages)
+                            .with_available_versions(&self.python_requirement, &self.visited, &self.index.packages)
                             .with_selector(self.selector.clone())
                             .with_python_requirement(&self.python_requirement)
                         )
@@ -504,6 +507,7 @@ impl<'a, Provider: ResolverProvider> Resolver<'a, Provider> {
                 // Wait for the metadata to be available.
                 let entry = self.index.packages.wait(package_name).await?;
                 let version_map = entry.value();
+                self.visited.insert(package_name.clone());
 
                 if let Some(extra) = extra {
                     debug!(
diff --git a/crates/puffin/tests/pip_install_scenarios.rs b/crates/puffin/tests/pip_install_scenarios.rs
index 563aa5c72..f62c65e51 100644
--- a/crates/puffin/tests/pip_install_scenarios.rs
+++ b/crates/puffin/tests/pip_install_scenarios.rs
@@ -566,8 +566,8 @@ fn dependency_excludes_range_of_compatible_versions() -> Result<()> {
 
               Because only albatross<=3.0.0 is available and albatross==3.0.0 depends on bluebird==3.0.0, we can conclude that albatross>=3.0.0 depends on bluebird==3.0.0.
               And because we know from (2) that all versions of crow, bluebird!=1.0.0, albatross<3.0.0 are incompatible, we can conclude that all versions of crow depend on one of:
-                  bluebird<=1.0.0
-                  bluebird>=3.0.0
+                  bluebird==1.0.0
+                  bluebird==3.0.0
 
               And because you require crow and you require bluebird>=2.0.0,<3.0.0, we can conclude that the requirements are unsatisfiable.
         "###);
@@ -697,8 +697,8 @@ fn dependency_excludes_non_contiguous_range_of_compatible_versions() -> Result<(
 
               Because only albatross<=3.0.0 is available and albatross==3.0.0 depends on bluebird==3.0.0, we can conclude that albatross>=3.0.0 depends on bluebird==3.0.0.
               And because we know from (2) that all versions of crow, bluebird!=1.0.0, albatross<3.0.0 are incompatible, we can conclude that all versions of crow depend on one of:
-                  bluebird<=1.0.0
-                  bluebird>=3.0.0
+                  bluebird==1.0.0
+                  bluebird==3.0.0
 
               And because you require crow and you require bluebird>=2.0.0,<3.0.0, we can conclude that the requirements are unsatisfiable.
         "###);