diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index aa2b23ec..4ee6a2fe 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -93,12 +93,25 @@ jobs:
             name_suffix: -gemma-e2b
             models: mlx-community/gemma-4-e2b-it-4bit
             test_model: mlx-community/gemma-4-e2b-it-4bit
+          # LFM2.5 used to be absent here and was fetched by the server mid-test, which
+          # is how it silently picked up a republished revision whose chat template was
+          # one brace short of valid and took main red. Prefetching it at a pinned
+          # revision is what makes this job depend on our code rather than on what
+          # upstream published that afternoon.
           - modality: vision
-            models: mlx-community/Qwen2-VL-2B-Instruct-4bit
+            models: mlx-community/Qwen2-VL-2B-Instruct-4bit LiquidAI/LFM2.5-VL-450M-MLX-4bit@10ce3604e42cd595497c47aaf67b7890e1e2a3b4
           - modality: audio
             models: mlx-community/gemma-4-e4b-it-4bit
           - modality: graph
             models: ""
+          # Synthetic checkpoints committed to the repository, a few hundred KB each,
+          # so this entry downloads nothing and needs no model cache. It covers the
+          # weight-and-config shapes behind #118, #120 and the b674 regression — the
+          # class of defect #128 found the unit suite has never caught, and which real
+          # checkpoints are too large to cover here (gemma-4-e2b alone is 3.6 GB
+          # against a 10 GB per-repository cache budget).
+          - modality: fixtures
+            models: ""
           # OpenAI-compatibility contract: what the client receives over the wire, once
           # real tokenisation decides chunk boundaries. Runs on the smallest model —
           # these assert server behaviour, not model quality (issue #128).
diff --git a/.github/workflows/update_dependencies.yml b/.github/workflows/update_dependencies.yml
index 66d02715..85a599c2 100644
--- a/.github/workflows/update_dependencies.yml
+++ b/.github/workflows/update_dependencies.yml
@@ -1,12 +1,25 @@
 name: Dependency Automation
 
+# Prepares a submodule bump when mlx-swift or mlx-swift-lm cuts a release, and stops
+# at a pushed branch rather than opening a pull request.
+#
+# Opening the PR from a workflow needs a personal access token, because GitHub does
+# not start workflow runs for events raised by GITHUB_TOKEN. A bot-opened PR would
+# therefore arrive with no checks at all — permanently pending, never green — and this
+# repository gates releases on CI concluding successfully (see release.yml). A branch
+# is the honest stopping point: opening the PR yourself takes one click, and CI then
+# runs normally because the event is yours.
+#
+# That a human sees the bump before it merges is a feature. Bumps here have needed a
+# pointer check, an umbrella build and a smoke test to be trustworthy; the automation
+# does the mechanical part and leaves the judgement.
+
 on:
   repository_dispatch:
     types: [dependency_bump]
 
 permissions:
   contents: write
-  pull-requests: write
 
 jobs:
   bump-dependencies:
@@ -18,34 +31,100 @@ jobs:
           submodules: recursive
           fetch-depth: 0
 
-      - name: Update swift-mlx dependencies
-        if: ${{ github.event.client_payload.source_repo == 'mlx-swift' }}
+      # client_payload is attacker-controlled in principle — anything able to dispatch
+      # to this repository chooses these strings — and they end up in shell and in a
+      # ref name. Validate them here and pass them onward through the environment
+      # rather than interpolating ${{ }} into a run block, where a crafted tag would
+      # be executed rather than compared.
+      - name: Validate dispatch payload
+        env:
+          PAYLOAD_SOURCE_REPO: ${{ github.event.client_payload.source_repo }}
+          PAYLOAD_NEW_TAG: ${{ github.event.client_payload.new_tag }}
         run: |
-          echo "Bumping mlx-swift dependency to ${{ github.event.client_payload.new_tag }}"
-          # In Package.swift we depend on branch main, but if we wanted to depend on a tag:
-          # SwiftPM resolves "main" to the latest commit automatically, but updating the SPM resolved file ensures deterministic builds:
-          swift package update mlx-swift
+          set -euo pipefail
+          case "$PAYLOAD_SOURCE_REPO" in
+            mlx-swift|mlx-swift-lm) ;;
+            *)
+              echo "::error::unexpected source_repo '$PAYLOAD_SOURCE_REPO' — expected mlx-swift or mlx-swift-lm"
+              exit 1
+              ;;
+          esac
+          if ! printf '%s' "$PAYLOAD_NEW_TAG" | grep -Eq '^[A-Za-z0-9][A-Za-z0-9._-]{0,63}$'; then
+            echo "::error::refusing tag '$PAYLOAD_NEW_TAG' — not a plain tag name"
+            exit 1
+          fi
+          {
+            echo "SOURCE_REPO=$PAYLOAD_SOURCE_REPO"
+            echo "NEW_TAG=$PAYLOAD_NEW_TAG"
+          } >> "$GITHUB_ENV"
 
-      - name: Update swift-mlx-lm dependencies
-        if: ${{ github.event.client_payload.source_repo == 'mlx-swift-lm' }}
+      # Both dependencies are `.package(path: "./…")` in Package.swift, backed by git
+      # submodules, so bumping either one is a pointer move. `swift package update`
+      # does nothing for a path dependency — SwiftPM takes whatever is on disk — which
+      # is why the mlx-swift branch of the previous version of this workflow could only
+      # ever have produced an empty commit.
+      - name: Move submodule to the released tag
         run: |
-          echo "Bumping local mlx-swift-lm submodule to ${{ github.event.client_payload.new_tag }}"
-          git submodule update --remote mlx-swift-lm
-          # Force the submodule onto the specific new release tag
-          cd mlx-swift-lm
-          git checkout ${{ github.event.client_payload.new_tag }}
-          cd ..
-          git add mlx-swift-lm
-
-      - name: Create Pull Request
-        uses: peter-evans/create-pull-request@v6
-        with:
-          token: ${{ secrets.SWIFTLM_PR_TOKEN }}
-          commit-message: "chore(deps): bump ${{ github.event.client_payload.source_repo }} to ${{ github.event.client_payload.new_tag }}"
-          title: "Update ${{ github.event.client_payload.source_repo }} Dependency to ${{ github.event.client_payload.new_tag }}"
-          body: |
-            Automated dependency update triggered by release `${{ github.event.client_payload.new_tag }}` in `SharpAI/${{ github.event.client_payload.source_repo }}`.
-            
-            This PR ensures SwiftLM is tracking the latest validated architectural improvements.
-          branch: "auto-update/${{ github.event.client_payload.source_repo }}-${{ github.event.client_payload.new_tag }}"
-          base: main
+          set -euo pipefail
+          git -C "$SOURCE_REPO" fetch --tags --force origin
+          if ! git -C "$SOURCE_REPO" rev-parse -q --verify "refs/tags/${NEW_TAG}^{commit}" >/dev/null; then
+            echo "::error::tag $NEW_TAG does not exist in $SOURCE_REPO"
+            exit 1
+          fi
+          before=$(git rev-parse "HEAD:$SOURCE_REPO")
+          git -C "$SOURCE_REPO" checkout --detach "refs/tags/$NEW_TAG"
+          after=$(git -C "$SOURCE_REPO" rev-parse HEAD)
+          {
+            echo "BEFORE_SHA=$before"
+            echo "AFTER_SHA=$after"
+          } >> "$GITHUB_ENV"
+          if [ "$before" = "$after" ]; then
+            echo "ALREADY_CURRENT=1" >> "$GITHUB_ENV"
+          fi
+
+      - name: Report an already-current submodule and stop
+        if: env.ALREADY_CURRENT == '1'
+        run: |
+          {
+            echo "### Nothing to bump"
+            echo
+            echo "\`$SOURCE_REPO\` is already at \`$NEW_TAG\` (\`${AFTER_SHA:0:7}\`)."
+          } >> "$GITHUB_STEP_SUMMARY"
+
+      - name: Push the bump branch
+        if: env.ALREADY_CURRENT != '1'
+        run: |
+          set -euo pipefail
+          branch="auto-update/${SOURCE_REPO}-${NEW_TAG}"
+          git checkout -B "$branch"
+          git add "$SOURCE_REPO"
+          git \
+            -c user.name='github-actions[bot]' \
+            -c user.email='41898282+github-actions[bot]@users.noreply.github.com' \
+            commit -m "chore(deps): bump $SOURCE_REPO to $NEW_TAG
+
+          Moves the $SOURCE_REPO submodule from ${BEFORE_SHA:0:7} to ${AFTER_SHA:0:7},
+          the commit tagged $NEW_TAG.
+
+          Prepared automatically; opened by hand so that CI runs against it."
+          # The auto-update/* namespace belongs to this workflow, so replacing a branch
+          # left by an earlier run for the same tag is safe and keeps re-runs idempotent.
+          git push --force origin "$branch"
+          echo "BUMP_BRANCH=$branch" >> "$GITHUB_ENV"
+
+      - name: Summarise, with a link that opens the pull request
+        if: env.ALREADY_CURRENT != '1'
+        run: |
+          {
+            echo "### \`$SOURCE_REPO\` → \`$NEW_TAG\` is ready"
+            echo
+            echo "Branch \`$BUMP_BRANCH\` pushed, moving the submodule from"
+            echo "\`${BEFORE_SHA:0:7}\` to \`${AFTER_SHA:0:7}\`."
+            echo
+            echo "**[Open the pull request](${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}/compare/main...${BUMP_BRANCH}?expand=1)**"
+            echo
+            echo "No PR is opened here on purpose: GitHub does not start workflow runs"
+            echo "for events raised by \`GITHUB_TOKEN\`, so a bot-opened PR would never"
+            echo "get CI. Opening it yourself gets the checks this repository gates"
+            echo "releases on."
+          } >> "$GITHUB_STEP_SUMMARY"
diff --git a/Sources/SwiftLM/Server.swift b/Sources/SwiftLM/Server.swift
index 626e4b80..f876a62c 100644
--- a/Sources/SwiftLM/Server.swift
+++ b/Sources/SwiftLM/Server.swift
@@ -43,15 +43,50 @@ private struct HubDownloader: Downloader, Sendable {
 }
 
 private struct TransformersTokenizerLoader: TokenizerLoader, Sendable {
+    /// Carried purely so template diagnostics can name the offending checkpoint; the
+    /// on-disk directory is a snapshot hash and means nothing to the person reading it.
+    let modelId: String
+    init(modelId: String = "this model") { self.modelId = modelId }
+
     func load(from directory: URL) async throws -> any MLXLMCommon.Tokenizer {
         let t = try await AutoTokenizer.from(modelFolder: directory)
-        return TransformersTokenizerBridge(t)
+        return TransformersTokenizerBridge(t, modelId: modelId)
+    }
+}
+
+/// A chat template that the Jinja parser rejected.
+///
+/// Worth its own type because the underlying error says only what the parser choked on
+/// — `parser('Unexpected token type: closeExpression')` — with no hint that the subject
+/// is a file shipped inside someone else's checkpoint. A model published with a
+/// malformed template loads fine, starts fine, and then fails every request, which
+/// reads as "the server is broken" rather than "this checkpoint is broken". That
+/// happened for real: LiquidAI republished LFM2.5-VL-450M-MLX-4bit with
+/// `{- bos_token -}}` (one brace short) and identifying it took a bisect across two
+/// model revisions.
+struct MalformedChatTemplate: Error, CustomStringConvertible {
+    let modelId: String
+    let underlying: String
+
+    var description: String {
+        """
+        The chat template shipped with \(modelId) is not valid Jinja and could not be \
+        parsed: \(underlying). This is a defect in the model's chat_template.jinja (or \
+        the chat_template field of its tokenizer_config.json), not in the request — \
+        report it to whoever publishes the checkpoint. Pinning to an earlier revision \
+        of the model is the usual workaround.
+        """
     }
 }
 
 private struct TransformersTokenizerBridge: MLXLMCommon.Tokenizer, Sendable {
     let upstream: any Tokenizers.Tokenizer
-    init(_ upstream: any Tokenizers.Tokenizer) { self.upstream = upstream }
+    /// Only used to name the model in template diagnostics.
+    let modelId: String
+    init(_ upstream: any Tokenizers.Tokenizer, modelId: String = "this model") {
+        self.upstream = upstream
+        self.modelId = modelId
+    }
     func encode(text: String, addSpecialTokens: Bool) -> [Int] {
         upstream.encode(text: text, addSpecialTokens: addSpecialTokens)
     }
@@ -73,6 +108,12 @@ private struct TransformersTokenizerBridge: MLXLMCommon.Tokenizer, Sendable {
                 messages: messages, tools: tools, additionalContext: additionalContext)
         } catch Tokenizers.TokenizerError.missingChatTemplate {
             throw MLXLMCommon.TokenizerError.missingChatTemplate
+        } catch {
+            // A missing template is a legitimate state (base models have none) and keeps
+            // its own error above. Anything else here means the template exists but the
+            // parser would not take it, which is worth naming precisely.
+            throw MalformedChatTemplate(
+                modelId: modelId, underlying: String(describing: error))
         }
     }
 }
@@ -598,7 +639,7 @@ struct MLXServer: AsyncParsableCommand {
             print("[SwiftLM] Loading Omni-Language Model (Text + Vision + Audio)...")
             container = try await OmniModelFactory.shared.loadContainer(
                 from: downloader,
-                using: TransformersTokenizerLoader(),
+                using: TransformersTokenizerLoader(modelId: resolvedModelId),
                 configuration: modelConfig
             ) { progress in
                 tracker.printProgress(progress)
@@ -607,7 +648,7 @@ struct MLXServer: AsyncParsableCommand {
             print("[SwiftLM] Loading VLM (vision-language model)...")
             container = try await VLMModelFactory.shared.loadContainer(
                 from: downloader,
-                using: TransformersTokenizerLoader(),
+                using: TransformersTokenizerLoader(modelId: resolvedModelId),
                 configuration: modelConfig
             ) { progress in
                 tracker.printProgress(progress)
@@ -618,7 +659,7 @@ struct MLXServer: AsyncParsableCommand {
             // and the native prepareForMultimodal path extracts real mel features.
             container = try await OmniModelFactory.shared.loadContainer(
                 from: downloader,
-                using: TransformersTokenizerLoader(),
+                using: TransformersTokenizerLoader(modelId: resolvedModelId),
                 configuration: modelConfig
             ) { progress in
                 tracker.printProgress(progress)
@@ -627,7 +668,7 @@ struct MLXServer: AsyncParsableCommand {
             print("[SwiftLM] Loading LLM (large language model)...")
             container = try await LLMModelFactory.shared.loadContainer(
                 from: downloader,
-                using: TransformersTokenizerLoader(),
+                using: TransformersTokenizerLoader(modelId: resolvedModelId),
                 configuration: modelConfig
             ) { progress in
                 tracker.printProgress(progress)
@@ -674,7 +715,7 @@ struct MLXServer: AsyncParsableCommand {
             let draftDownloader = HubDownloader(hub: HubApi(downloadBase: cacheRoot))
             let draftContainer = try await LLMModelFactory.shared.loadContainer(
                 from: draftDownloader,
-                using: TransformersTokenizerLoader(),
+                using: TransformersTokenizerLoader(modelId: resolvedModelId),
                 configuration: draftConfig
             ) { progress in
                 // Silent loading for draft model
@@ -709,7 +750,7 @@ struct MLXServer: AsyncParsableCommand {
             let assistantDownloader = HubDownloader(hub: HubApi(downloadBase: cacheRoot))
             let assistantContainer = try await LLMModelFactory.shared.loadContainer(
                 from: assistantDownloader,
-                using: TransformersTokenizerLoader(),
+                using: TransformersTokenizerLoader(modelId: resolvedModelId),
                 configuration: assistantConfig
             ) { _ in }
             mtpAssistantModelRef = await assistantContainer.perform { assistantContext in
@@ -837,6 +878,34 @@ struct MLXServer: AsyncParsableCommand {
             print("[SwiftLM] 🧠 Auto-calibration (Wisdom) bypassed for SSD Streaming")
         }
 
+        // Render the chat template once before opening the port. A checkpoint whose
+        // template does not parse otherwise loads cleanly, reports ready, and then fails
+        // every single request — a shape that reads as a broken server rather than a
+        // broken model, and that survives restarts without ever explaining itself.
+        // Failing here instead costs one template render and makes the cause the first
+        // thing anybody sees.
+        //
+        // A model with no chat template at all is a different and legitimate case (base
+        // models ship without one, and /v1/completions does not need it), so that is
+        // passed over rather than treated as a defect.
+        do {
+            let probeTokenizer = await container.tokenizer
+            // Shaped like the simplest real request rather than a bare minimum: one user
+            // turn with a generation prompt is what every chat template is written to
+            // handle, so a failure here is the template's, not the probe's.
+            _ = try probeTokenizer.applyChatTemplate(
+                messages: [["role": "user", "content": "ping"]],
+                tools: nil,
+                additionalContext: ["add_generation_prompt": true]
+            )
+        } catch let error as MalformedChatTemplate {
+            print("[SwiftLM] ❌ \(error.description)")
+            throw error
+        } catch {
+            // Missing template, or a template that needs context this probe does not
+            // supply. Neither is a reason to refuse to start.
+        }
+
         print("[SwiftLM] Model loaded. Starting HTTP server on \(host):\(port)")
 
         // ── Capture CLI defaults into a shared config ──
diff --git a/mlx-swift-lm b/mlx-swift-lm
index 6a2c1799..bfc2462b 160000
--- a/mlx-swift-lm
+++ b/mlx-swift-lm
@@ -1 +1 @@
-Subproject commit 6a2c179998723107bb3d271963eeb9061056bee5
+Subproject commit bfc2462b975bc278411be041d5761299e836c846
diff --git a/scripts/ci-download-models.sh b/scripts/ci-download-models.sh
index a2fa868a..e0632346 100755
--- a/scripts/ci-download-models.sh
+++ b/scripts/ci-download-models.sh
@@ -30,12 +30,26 @@ has_partial_files() {
 }
 
 download_one() {
-    local repo="$1"
+    local spec="$1"
+    # `repo@revision` pins to an immutable commit. Upstream repositories are mutable:
+    # LiquidAI republished LFM2.5-VL-450M-MLX-4bit with a malformed chat template
+    # (`{- bos_token -}}`, one brace short), which turned every request into an HTTP 500
+    # and took main red with no change on our side. A floating tag means CI results
+    # depend on what a third party did that afternoon.
+    local repo="${spec%@*}"
+    local revision=""
+    if [ "$spec" != "$repo" ]; then
+        revision="${spec##*@}"
+    fi
     local dir="$HUB_DIR/models--${repo//\//--}"
 
     for attempt in $(seq 1 "$ATTEMPTS"); do
-        echo "--- $repo (attempt $attempt/$ATTEMPTS, per-request timeout ${HF_HUB_DOWNLOAD_TIMEOUT}s)"
-        if hf download "$repo"; then
+        echo "--- $repo${revision:+ @ $revision} (attempt $attempt/$ATTEMPTS, per-request timeout ${HF_HUB_DOWNLOAD_TIMEOUT}s)"
+        # Spelled out rather than assembling an args array: the runners are macOS, which
+        # ships bash 3.2, where expanding an empty array under `set -u` is an unbound
+        # variable error rather than nothing at all. That fails every unpinned download.
+        if { [ -n "$revision" ] && hf download "$repo" --revision "$revision"; } \
+            || { [ -z "$revision" ] && hf download "$repo"; }; then
             if has_partial_files "$dir"; then
                 echo "::warning::$repo downloaded but .incomplete files remain; retrying"
             else
diff --git a/scripts/make-test-fixtures.py b/scripts/make-test-fixtures.py
new file mode 100755
index 00000000..e61ecf35
--- /dev/null
+++ b/scripts/make-test-fixtures.py
@@ -0,0 +1,250 @@
+#!/usr/bin/env python3
+"""Generate tiny synthetic checkpoints that exercise the real model-load path.
+
+Issue #128 observed that every defect in the #108/#110/#112 cycle was caught either
+by loading a real checkpoint or by code review, and none by the unit suite — because
+the bugs lived in weight-and-config-shape assumptions that only a checkpoint on disk
+exercises. The obvious fix, running CI against real models, runs into arithmetic:
+gemma-4-e2b is 3.6 GB and GitHub allows 10 GB of cache for the whole repository.
+
+llama.cpp solved the same problem by publishing purpose-built tiny models
+(`ggml-org/test-model-stories260K`, 1.2 MB) rather than shrinking real ones. Their
+files are GGUF and unusable here, but the technique transfers: a checkpoint with the
+same *shape* as a real one — same config fields, same weight keys, random values,
+a ~300-token vocabulary instead of 150k — runs the same loading code and weighs a
+few hundred kilobytes.
+
+What these fixtures test is the plumbing: config parsing, weight-key resolution,
+sanitisation, layer materialisation, quantisation metadata. They say nothing about
+whether the arithmetic is correct, because the weights are noise. Real checkpoints
+remain the only way to judge output quality.
+
+Usage:  python3 scripts/make-test-fixtures.py [output-dir]
+"""
+import json
+import os
+import sys
+
+import numpy as np
+from safetensors.numpy import save_file
+from tokenizers import Tokenizer, decoders, models, pre_tokenizers, processors
+
+ROOT = sys.argv[1] if len(sys.argv) > 1 else "tests/fixtures"
+
+VOCAB = 288
+SPECIALS = ["<|endoftext|>", "<|im_start|>", "<|im_end|>", "<pad>"]
+CHAT_TEMPLATE = (
+    "{% for m in messages %}<|im_start|>{{ m['role'] }}\n{{ m['content'] }}<|im_end|>\n"
+    "{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"
+)
+
+
+def write_tokenizer(out):
+    """A byte-level BPE, which is what Qwen/GPT-2 ship.
+
+    Not a WordLevel model: swift-transformers rejects those with "BPETokenizer
+    requires merges". Merges must also be spelled in the byte-level alphabet — "Ġ"
+    for a space rather than a raw 0x20 — or the tokenizers library refuses to build.
+    """
+    vocab = {t: i for i, t in enumerate(SPECIALS)}
+    for ch in sorted(pre_tokenizers.ByteLevel.alphabet()):
+        vocab[ch] = len(vocab)
+    merges = [("h", "e"), ("l", "l"), ("he", "ll"), ("t", "e")]
+    for a, b in merges:
+        vocab.setdefault(a + b, len(vocab))
+    while len(vocab) < VOCAB:
+        vocab[f"<|unused{len(vocab)}|>"] = len(vocab)
+
+    tok = Tokenizer(models.BPE(vocab=vocab, merges=merges, unk_token=None, fuse_unk=False))
+    tok.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=False)
+    tok.decoder = decoders.ByteLevel()
+    tok.post_processor = processors.ByteLevel(trim_offsets=False)
+    tok.add_special_tokens(SPECIALS)
+    tok.save(os.path.join(out, "tokenizer.json"))
+
+    json.dump(
+        {
+            "tokenizer_class": "PreTrainedTokenizerFast",
+            "bos_token": "<|endoftext|>",
+            "eos_token": "<|im_end|>",
+            "pad_token": "<pad>",
+            "unk_token": "<|endoftext|>",
+            "chat_template": CHAT_TEMPLATE,
+        },
+        open(os.path.join(out, "tokenizer_config.json"), "w"),
+        indent=2,
+    )
+
+
+def rand(rng, *shape):
+    return (rng.standard_normal(shape) * 0.02).astype(np.float16)
+
+
+def ones(n):
+    return np.ones(n, np.float16)
+
+
+def build_dense(out, stray_shard=False):
+    """A plain Qwen2. With stray_shard, an extra .safetensors sits beside the real one
+    without appearing in the weight index — the shape behind #118, where files outside
+    the index were being loaded and their keys rejected."""
+    H, L, HEADS, KVH, INTER = 64, 2, 4, 2, 128
+    HD = H // HEADS
+    rng = np.random.default_rng(0)
+
+    json.dump(
+        {
+            "model_type": "qwen2",
+            "architectures": ["Qwen2ForCausalLM"],
+            "vocab_size": VOCAB,
+            "hidden_size": H,
+            "intermediate_size": INTER,
+            "num_hidden_layers": L,
+            "num_attention_heads": HEADS,
+            "num_key_value_heads": KVH,
+            "max_position_embeddings": 512,
+            "rms_norm_eps": 1e-6,
+            "rope_theta": 10000.0,
+            "tie_word_embeddings": False,
+        },
+        open(os.path.join(out, "config.json"), "w"),
+        indent=2,
+    )
+
+    w = {
+        "model.embed_tokens.weight": rand(rng, VOCAB, H),
+        "model.norm.weight": ones(H),
+        "lm_head.weight": rand(rng, VOCAB, H),
+    }
+    for i in range(L):
+        p = f"model.layers.{i}"
+        w[f"{p}.self_attn.q_proj.weight"] = rand(rng, HEADS * HD, H)
+        w[f"{p}.self_attn.q_proj.bias"] = rand(rng, HEADS * HD)
+        w[f"{p}.self_attn.k_proj.weight"] = rand(rng, KVH * HD, H)
+        w[f"{p}.self_attn.k_proj.bias"] = rand(rng, KVH * HD)
+        w[f"{p}.self_attn.v_proj.weight"] = rand(rng, KVH * HD, H)
+        w[f"{p}.self_attn.v_proj.bias"] = rand(rng, KVH * HD)
+        w[f"{p}.self_attn.o_proj.weight"] = rand(rng, H, HEADS * HD)
+        w[f"{p}.mlp.gate_proj.weight"] = rand(rng, INTER, H)
+        w[f"{p}.mlp.up_proj.weight"] = rand(rng, INTER, H)
+        w[f"{p}.mlp.down_proj.weight"] = rand(rng, H, INTER)
+        w[f"{p}.input_layernorm.weight"] = ones(H)
+        w[f"{p}.post_attention_layernorm.weight"] = ones(H)
+
+    save_file(w, os.path.join(out, "model.safetensors"), metadata={"format": "pt"})
+
+    if stray_shard:
+        # Deliberately absent from weight_map. A loader that globs *.safetensors instead
+        # of reading the index picks this up and fails on the unknown key.
+        save_file(
+            {"not_a_real_module.weight": rand(rng, 8, 8)},
+            os.path.join(out, "extra-not-in-index.safetensors"),
+            metadata={"format": "pt"},
+        )
+        json.dump(
+            {
+                "metadata": {"total_size": 0},
+                "weight_map": {k: "model.safetensors" for k in w},
+            },
+            open(os.path.join(out, "model.safetensors.index.json"), "w"),
+        )
+    return len(w)
+
+
+def build_gemma4_kv_shared(out, vestigial):
+    """Gemma 4 text with KV-shared layers.
+
+    Two real checkpoints disagree about what a shared layer ships: gemma-4-e4b omits
+    its k/v projections, gemma-4-e2b includes them anyway. #120 was the first case
+    failing to load; the b674 regression was the second, after a fix that assumed the
+    first was universal. `vestigial=True` is the e2b shape.
+
+    Shapes mirror a real gemma-4-e2b checkpoint rather than being guessed.
+    """
+    H, L, SHARED, HEADS, KVH, HD = 64, 4, 2, 4, 2, 16
+    INTER, PLI, VPLI = 128, 32, 16
+    rng = np.random.default_rng(0)
+
+    json.dump(
+        {
+            "model_type": "gemma4_text",
+            "architectures": ["Gemma4ForCausalLM"],
+            "hidden_size": H,
+            "num_hidden_layers": L,
+            "intermediate_size": INTER,
+            "num_attention_heads": HEADS,
+            "head_dim": HD,
+            "global_head_dim": HD,
+            "rms_norm_eps": 1e-6,
+            "vocab_size": VOCAB,
+            "num_key_value_heads": KVH,
+            "rope_traditional": False,
+            "rope_theta": 10000.0,
+            "sliding_window": 128,
+            "sliding_window_pattern": 1,
+            "max_position_embeddings": 512,
+            "num_kv_shared_layers": SHARED,
+            "use_double_wide_mlp": False,
+            "tie_word_embeddings": True,
+            "hidden_size_per_layer_input": PLI,
+            "vocab_size_per_layer_input": VPLI,
+            "final_logit_softcapping": 30.0,
+            "enable_moe_block": False,
+            "attention_k_eq_v": False,
+        },
+        open(os.path.join(out, "config.json"), "w"),
+        indent=2,
+    )
+
+    w = {
+        "model.embed_tokens.weight": rand(rng, VOCAB, H),
+        "model.norm.weight": ones(H),
+        "model.embed_tokens_per_layer.weight": rand(rng, VPLI, L * PLI),
+        "model.per_layer_model_projection.weight": rand(rng, L * PLI, H),
+        "model.per_layer_projection_norm.weight": ones(PLI),
+    }
+    boundary = L - SHARED
+    for i in range(L):
+        p = f"model.layers.{i}"
+        w[f"{p}.self_attn.q_proj.weight"] = rand(rng, HEADS * HD, H)
+        w[f"{p}.self_attn.o_proj.weight"] = rand(rng, H, HEADS * HD)
+        w[f"{p}.self_attn.q_norm.weight"] = ones(HD)
+        w[f"{p}.layer_scalar"] = np.ones(1, np.float16)
+        if i < boundary or vestigial:
+            w[f"{p}.self_attn.k_proj.weight"] = rand(rng, KVH * HD, H)
+            w[f"{p}.self_attn.v_proj.weight"] = rand(rng, KVH * HD, H)
+            w[f"{p}.self_attn.k_norm.weight"] = ones(HD)
+        w[f"{p}.mlp.gate_proj.weight"] = rand(rng, INTER, H)
+        w[f"{p}.mlp.up_proj.weight"] = rand(rng, INTER, H)
+        w[f"{p}.mlp.down_proj.weight"] = rand(rng, H, INTER)
+        w[f"{p}.input_layernorm.weight"] = ones(H)
+        w[f"{p}.post_attention_layernorm.weight"] = ones(H)
+        w[f"{p}.pre_feedforward_layernorm.weight"] = ones(H)
+        w[f"{p}.post_feedforward_layernorm.weight"] = ones(H)
+        w[f"{p}.per_layer_input_gate.weight"] = rand(rng, PLI, H)
+        w[f"{p}.per_layer_projection.weight"] = rand(rng, H, PLI)
+        w[f"{p}.post_per_layer_input_norm.weight"] = ones(H)
+
+    save_file(w, os.path.join(out, "model.safetensors"), metadata={"format": "pt"})
+    return len(w)
+
+
+FIXTURES = {
+    "dense": (build_dense, {}),
+    "stray-shard": (build_dense, {"stray_shard": True}),
+    "kv-shared-absent": (build_gemma4_kv_shared, {"vestigial": False}),
+    "kv-shared-present": (build_gemma4_kv_shared, {"vestigial": True}),
+}
+
+if __name__ == "__main__":
+    os.makedirs(ROOT, exist_ok=True)
+    total = 0
+    for name, (fn, kwargs) in FIXTURES.items():
+        out = os.path.join(ROOT, name)
+        os.makedirs(out, exist_ok=True)
+        n = fn(out, **kwargs)
+        write_tokenizer(out)
+        size = sum(os.path.getsize(os.path.join(out, f)) for f in os.listdir(out))
+        total += size
+        print(f"  {name:<20} {n:>3} tensors  {size/1024:>7.1f} KB")
+    print(f"  {'total':<20} {'':>3}          {total/1024:>7.1f} KB")
diff --git a/tests/fixtures/dense/config.json b/tests/fixtures/dense/config.json
new file mode 100644
index 00000000..5f3594c0
--- /dev/null
+++ b/tests/fixtures/dense/config.json
@@ -0,0 +1,16 @@
+{
+  "model_type": "qwen2",
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "vocab_size": 288,
+  "hidden_size": 64,
+  "intermediate_size": 128,
+  "num_hidden_layers": 2,
+  "num_attention_heads": 4,
+  "num_key_value_heads": 2,
+  "max_position_embeddings": 512,
+  "rms_norm_eps": 1e-06,
+  "rope_theta": 10000.0,
+  "tie_word_embeddings": false
+}
\ No newline at end of file
diff --git a/tests/fixtures/dense/model.safetensors b/tests/fixtures/dense/model.safetensors
new file mode 100644
index 00000000..6d98f758
Binary files /dev/null and b/tests/fixtures/dense/model.safetensors differ
diff --git a/tests/fixtures/dense/tokenizer.json b/tests/fixtures/dense/tokenizer.json
new file mode 100644
index 00000000..21c0e134
--- /dev/null
+++ b/tests/fixtures/dense/tokenizer.json
@@ -0,0 +1,380 @@
+{
+  "version": "1.0",
+  "truncation": null,
+  "padding": null,
+  "added_tokens": [
+    {
+      "id": 0,
+      "content": "<|endoftext|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 1,
+      "content": "<|im_start|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 2,
+      "content": "<|im_end|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 3,
+      "content": "<pad>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    }
+  ],
+  "normalizer": null,
+  "pre_tokenizer": {
+    "type": "ByteLevel",
+    "add_prefix_space": false,
+    "trim_offsets": true,
+    "use_regex": true
+  },
+  "post_processor": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": false,
+    "use_regex": true
+  },
+  "decoder": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": true,
+    "use_regex": true
+  },
+  "model": {
+    "type": "BPE",
+    "dropout": null,
+    "unk_token": null,
+    "continuing_subword_prefix": null,
+    "end_of_word_suffix": null,
+    "fuse_unk": false,
+    "byte_fallback": false,
+    "ignore_merges": false,
+    "vocab": {
+      "<|endoftext|>": 0,
+      "<|im_start|>": 1,
+      "<|im_end|>": 2,
+      "<pad>": 3,
+      "!": 4,
+      "\"": 5,
+      "#": 6,
+      "$": 7,
+      "%": 8,
+      "&": 9,
+      "'": 10,
+      "(": 11,
+      ")": 12,
+      "*": 13,
+      "+": 14,
+      ",": 15,
+      "-": 16,
+      ".": 17,
+      "/": 18,
+      "0": 19,
+      "1": 20,
+      "2": 21,
+      "3": 22,
+      "4": 23,
+      "5": 24,
+      "6": 25,
+      "7": 26,
+      "8": 27,
+      "9": 28,
+      ":": 29,
+      ";": 30,
+      "<": 31,
+      "=": 32,
+      ">": 33,
+      "?": 34,
+      "@": 35,
+      "A": 36,
+      "B": 37,
+      "C": 38,
+      "D": 39,
+      "E": 40,
+      "F": 41,
+      "G": 42,
+      "H": 43,
+      "I": 44,
+      "J": 45,
+      "K": 46,
+      "L": 47,
+      "M": 48,
+      "N": 49,
+      "O": 50,
+      "P": 51,
+      "Q": 52,
+      "R": 53,
+      "S": 54,
+      "T": 55,
+      "U": 56,
+      "V": 57,
+      "W": 58,
+      "X": 59,
+      "Y": 60,
+      "Z": 61,
+      "[": 62,
+      "\\": 63,
+      "]": 64,
+      "^": 65,
+      "_": 66,
+      "`": 67,
+      "a": 68,
+      "b": 69,
+      "c": 70,
+      "d": 71,
+      "e": 72,
+      "f": 73,
+      "g": 74,
+      "h": 75,
+      "i": 76,
+      "j": 77,
+      "k": 78,
+      "l": 79,
+      "m": 80,
+      "n": 81,
+      "o": 82,
+      "p": 83,
+      "q": 84,
+      "r": 85,
+      "s": 86,
+      "t": 87,
+      "u": 88,
+      "v": 89,
+      "w": 90,
+      "x": 91,
+      "y": 92,
+      "z": 93,
+      "{": 94,
+      "|": 95,
+      "}": 96,
+      "~": 97,
+      "¡": 98,
+      "¢": 99,
+      "£": 100,
+      "¤": 101,
+      "¥": 102,
+      "¦": 103,
+      "§": 104,
+      "¨": 105,
+      "©": 106,
+      "ª": 107,
+      "«": 108,
+      "¬": 109,
+      "®": 110,
+      "¯": 111,
+      "°": 112,
+      "±": 113,
+      "²": 114,
+      "³": 115,
+      "´": 116,
+      "µ": 117,
+      "¶": 118,
+      "·": 119,
+      "¸": 120,
+      "¹": 121,
+      "º": 122,
+      "»": 123,
+      "¼": 124,
+      "½": 125,
+      "¾": 126,
+      "¿": 127,
+      "À": 128,
+      "Á": 129,
+      "Â": 130,
+      "Ã": 131,
+      "Ä": 132,
+      "Å": 133,
+      "Æ": 134,
+      "Ç": 135,
+      "È": 136,
+      "É": 137,
+      "Ê": 138,
+      "Ë": 139,
+      "Ì": 140,
+      "Í": 141,
+      "Î": 142,
+      "Ï": 143,
+      "Ð": 144,
+      "Ñ": 145,
+      "Ò": 146,
+      "Ó": 147,
+      "Ô": 148,
+      "Õ": 149,
+      "Ö": 150,
+      "×": 151,
+      "Ø": 152,
+      "Ù": 153,
+      "Ú": 154,
+      "Û": 155,
+      "Ü": 156,
+      "Ý": 157,
+      "Þ": 158,
+      "ß": 159,
+      "à": 160,
+      "á": 161,
+      "â": 162,
+      "ã": 163,
+      "ä": 164,
+      "å": 165,
+      "æ": 166,
+      "ç": 167,
+      "è": 168,
+      "é": 169,
+      "ê": 170,
+      "ë": 171,
+      "ì": 172,
+      "í": 173,
+      "î": 174,
+      "ï": 175,
+      "ð": 176,
+      "ñ": 177,
+      "ò": 178,
+      "ó": 179,
+      "ô": 180,
+      "õ": 181,
+      "ö": 182,
+      "÷": 183,
+      "ø": 184,
+      "ù": 185,
+      "ú": 186,
+      "û": 187,
+      "ü": 188,
+      "ý": 189,
+      "þ": 190,
+      "ÿ": 191,
+      "Ā": 192,
+      "ā": 193,
+      "Ă": 194,
+      "ă": 195,
+      "Ą": 196,
+      "ą": 197,
+      "Ć": 198,
+      "ć": 199,
+      "Ĉ": 200,
+      "ĉ": 201,
+      "Ċ": 202,
+      "ċ": 203,
+      "Č": 204,
+      "č": 205,
+      "Ď": 206,
+      "ď": 207,
+      "Đ": 208,
+      "đ": 209,
+      "Ē": 210,
+      "ē": 211,
+      "Ĕ": 212,
+      "ĕ": 213,
+      "Ė": 214,
+      "ė": 215,
+      "Ę": 216,
+      "ę": 217,
+      "Ě": 218,
+      "ě": 219,
+      "Ĝ": 220,
+      "ĝ": 221,
+      "Ğ": 222,
+      "ğ": 223,
+      "Ġ": 224,
+      "ġ": 225,
+      "Ģ": 226,
+      "ģ": 227,
+      "Ĥ": 228,
+      "ĥ": 229,
+      "Ħ": 230,
+      "ħ": 231,
+      "Ĩ": 232,
+      "ĩ": 233,
+      "Ī": 234,
+      "ī": 235,
+      "Ĭ": 236,
+      "ĭ": 237,
+      "Į": 238,
+      "į": 239,
+      "İ": 240,
+      "ı": 241,
+      "IJ": 242,
+      "ij": 243,
+      "Ĵ": 244,
+      "ĵ": 245,
+      "Ķ": 246,
+      "ķ": 247,
+      "ĸ": 248,
+      "Ĺ": 249,
+      "ĺ": 250,
+      "Ļ": 251,
+      "ļ": 252,
+      "Ľ": 253,
+      "ľ": 254,
+      "Ŀ": 255,
+      "ŀ": 256,
+      "Ł": 257,
+      "ł": 258,
+      "Ń": 259,
+      "he": 260,
+      "ll": 261,
+      "hell": 262,
+      "te": 263,
+      "<|unused264|>": 264,
+      "<|unused265|>": 265,
+      "<|unused266|>": 266,
+      "<|unused267|>": 267,
+      "<|unused268|>": 268,
+      "<|unused269|>": 269,
+      "<|unused270|>": 270,
+      "<|unused271|>": 271,
+      "<|unused272|>": 272,
+      "<|unused273|>": 273,
+      "<|unused274|>": 274,
+      "<|unused275|>": 275,
+      "<|unused276|>": 276,
+      "<|unused277|>": 277,
+      "<|unused278|>": 278,
+      "<|unused279|>": 279,
+      "<|unused280|>": 280,
+      "<|unused281|>": 281,
+      "<|unused282|>": 282,
+      "<|unused283|>": 283,
+      "<|unused284|>": 284,
+      "<|unused285|>": 285,
+      "<|unused286|>": 286,
+      "<|unused287|>": 287
+    },
+    "merges": [
+      [
+        "h",
+        "e"
+      ],
+      [
+        "l",
+        "l"
+      ],
+      [
+        "he",
+        "ll"
+      ],
+      [
+        "t",
+        "e"
+      ]
+    ]
+  }
+}
\ No newline at end of file
diff --git a/tests/fixtures/dense/tokenizer_config.json b/tests/fixtures/dense/tokenizer_config.json
new file mode 100644
index 00000000..ee42569a
--- /dev/null
+++ b/tests/fixtures/dense/tokenizer_config.json
@@ -0,0 +1,8 @@
+{
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|im_end|>",
+  "pad_token": "<pad>",
+  "unk_token": "<|endoftext|>",
+  "chat_template": "{% for m in messages %}<|im_start|>{{ m['role'] }}\n{{ m['content'] }}<|im_end|>\n{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"
+}
\ No newline at end of file
diff --git a/tests/fixtures/kv-shared-absent/config.json b/tests/fixtures/kv-shared-absent/config.json
new file mode 100644
index 00000000..78cfc80b
--- /dev/null
+++ b/tests/fixtures/kv-shared-absent/config.json
@@ -0,0 +1,28 @@
+{
+  "model_type": "gemma4_text",
+  "architectures": [
+    "Gemma4ForCausalLM"
+  ],
+  "hidden_size": 64,
+  "num_hidden_layers": 4,
+  "intermediate_size": 128,
+  "num_attention_heads": 4,
+  "head_dim": 16,
+  "global_head_dim": 16,
+  "rms_norm_eps": 1e-06,
+  "vocab_size": 288,
+  "num_key_value_heads": 2,
+  "rope_traditional": false,
+  "rope_theta": 10000.0,
+  "sliding_window": 128,
+  "sliding_window_pattern": 1,
+  "max_position_embeddings": 512,
+  "num_kv_shared_layers": 2,
+  "use_double_wide_mlp": false,
+  "tie_word_embeddings": true,
+  "hidden_size_per_layer_input": 32,
+  "vocab_size_per_layer_input": 16,
+  "final_logit_softcapping": 30.0,
+  "enable_moe_block": false,
+  "attention_k_eq_v": false
+}
\ No newline at end of file
diff --git a/tests/fixtures/kv-shared-absent/model.safetensors b/tests/fixtures/kv-shared-absent/model.safetensors
new file mode 100644
index 00000000..6515c0f4
Binary files /dev/null and b/tests/fixtures/kv-shared-absent/model.safetensors differ
diff --git a/tests/fixtures/kv-shared-absent/tokenizer.json b/tests/fixtures/kv-shared-absent/tokenizer.json
new file mode 100644
index 00000000..21c0e134
--- /dev/null
+++ b/tests/fixtures/kv-shared-absent/tokenizer.json
@@ -0,0 +1,380 @@
+{
+  "version": "1.0",
+  "truncation": null,
+  "padding": null,
+  "added_tokens": [
+    {
+      "id": 0,
+      "content": "<|endoftext|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 1,
+      "content": "<|im_start|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 2,
+      "content": "<|im_end|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 3,
+      "content": "<pad>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    }
+  ],
+  "normalizer": null,
+  "pre_tokenizer": {
+    "type": "ByteLevel",
+    "add_prefix_space": false,
+    "trim_offsets": true,
+    "use_regex": true
+  },
+  "post_processor": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": false,
+    "use_regex": true
+  },
+  "decoder": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": true,
+    "use_regex": true
+  },
+  "model": {
+    "type": "BPE",
+    "dropout": null,
+    "unk_token": null,
+    "continuing_subword_prefix": null,
+    "end_of_word_suffix": null,
+    "fuse_unk": false,
+    "byte_fallback": false,
+    "ignore_merges": false,
+    "vocab": {
+      "<|endoftext|>": 0,
+      "<|im_start|>": 1,
+      "<|im_end|>": 2,
+      "<pad>": 3,
+      "!": 4,
+      "\"": 5,
+      "#": 6,
+      "$": 7,
+      "%": 8,
+      "&": 9,
+      "'": 10,
+      "(": 11,
+      ")": 12,
+      "*": 13,
+      "+": 14,
+      ",": 15,
+      "-": 16,
+      ".": 17,
+      "/": 18,
+      "0": 19,
+      "1": 20,
+      "2": 21,
+      "3": 22,
+      "4": 23,
+      "5": 24,
+      "6": 25,
+      "7": 26,
+      "8": 27,
+      "9": 28,
+      ":": 29,
+      ";": 30,
+      "<": 31,
+      "=": 32,
+      ">": 33,
+      "?": 34,
+      "@": 35,
+      "A": 36,
+      "B": 37,
+      "C": 38,
+      "D": 39,
+      "E": 40,
+      "F": 41,
+      "G": 42,
+      "H": 43,
+      "I": 44,
+      "J": 45,
+      "K": 46,
+      "L": 47,
+      "M": 48,
+      "N": 49,
+      "O": 50,
+      "P": 51,
+      "Q": 52,
+      "R": 53,
+      "S": 54,
+      "T": 55,
+      "U": 56,
+      "V": 57,
+      "W": 58,
+      "X": 59,
+      "Y": 60,
+      "Z": 61,
+      "[": 62,
+      "\\": 63,
+      "]": 64,
+      "^": 65,
+      "_": 66,
+      "`": 67,
+      "a": 68,
+      "b": 69,
+      "c": 70,
+      "d": 71,
+      "e": 72,
+      "f": 73,
+      "g": 74,
+      "h": 75,
+      "i": 76,
+      "j": 77,
+      "k": 78,
+      "l": 79,
+      "m": 80,
+      "n": 81,
+      "o": 82,
+      "p": 83,
+      "q": 84,
+      "r": 85,
+      "s": 86,
+      "t": 87,
+      "u": 88,
+      "v": 89,
+      "w": 90,
+      "x": 91,
+      "y": 92,
+      "z": 93,
+      "{": 94,
+      "|": 95,
+      "}": 96,
+      "~": 97,
+      "¡": 98,
+      "¢": 99,
+      "£": 100,
+      "¤": 101,
+      "¥": 102,
+      "¦": 103,
+      "§": 104,
+      "¨": 105,
+      "©": 106,
+      "ª": 107,
+      "«": 108,
+      "¬": 109,
+      "®": 110,
+      "¯": 111,
+      "°": 112,
+      "±": 113,
+      "²": 114,
+      "³": 115,
+      "´": 116,
+      "µ": 117,
+      "¶": 118,
+      "·": 119,
+      "¸": 120,
+      "¹": 121,
+      "º": 122,
+      "»": 123,
+      "¼": 124,
+      "½": 125,
+      "¾": 126,
+      "¿": 127,
+      "À": 128,
+      "Á": 129,
+      "Â": 130,
+      "Ã": 131,
+      "Ä": 132,
+      "Å": 133,
+      "Æ": 134,
+      "Ç": 135,
+      "È": 136,
+      "É": 137,
+      "Ê": 138,
+      "Ë": 139,
+      "Ì": 140,
+      "Í": 141,
+      "Î": 142,
+      "Ï": 143,
+      "Ð": 144,
+      "Ñ": 145,
+      "Ò": 146,
+      "Ó": 147,
+      "Ô": 148,
+      "Õ": 149,
+      "Ö": 150,
+      "×": 151,
+      "Ø": 152,
+      "Ù": 153,
+      "Ú": 154,
+      "Û": 155,
+      "Ü": 156,
+      "Ý": 157,
+      "Þ": 158,
+      "ß": 159,
+      "à": 160,
+      "á": 161,
+      "â": 162,
+      "ã": 163,
+      "ä": 164,
+      "å": 165,
+      "æ": 166,
+      "ç": 167,
+      "è": 168,
+      "é": 169,
+      "ê": 170,
+      "ë": 171,
+      "ì": 172,
+      "í": 173,
+      "î": 174,
+      "ï": 175,
+      "ð": 176,
+      "ñ": 177,
+      "ò": 178,
+      "ó": 179,
+      "ô": 180,
+      "õ": 181,
+      "ö": 182,
+      "÷": 183,
+      "ø": 184,
+      "ù": 185,
+      "ú": 186,
+      "û": 187,
+      "ü": 188,
+      "ý": 189,
+      "þ": 190,
+      "ÿ": 191,
+      "Ā": 192,
+      "ā": 193,
+      "Ă": 194,
+      "ă": 195,
+      "Ą": 196,
+      "ą": 197,
+      "Ć": 198,
+      "ć": 199,
+      "Ĉ": 200,
+      "ĉ": 201,
+      "Ċ": 202,
+      "ċ": 203,
+      "Č": 204,
+      "č": 205,
+      "Ď": 206,
+      "ď": 207,
+      "Đ": 208,
+      "đ": 209,
+      "Ē": 210,
+      "ē": 211,
+      "Ĕ": 212,
+      "ĕ": 213,
+      "Ė": 214,
+      "ė": 215,
+      "Ę": 216,
+      "ę": 217,
+      "Ě": 218,
+      "ě": 219,
+      "Ĝ": 220,
+      "ĝ": 221,
+      "Ğ": 222,
+      "ğ": 223,
+      "Ġ": 224,
+      "ġ": 225,
+      "Ģ": 226,
+      "ģ": 227,
+      "Ĥ": 228,
+      "ĥ": 229,
+      "Ħ": 230,
+      "ħ": 231,
+      "Ĩ": 232,
+      "ĩ": 233,
+      "Ī": 234,
+      "ī": 235,
+      "Ĭ": 236,
+      "ĭ": 237,
+      "Į": 238,
+      "į": 239,
+      "İ": 240,
+      "ı": 241,
+      "IJ": 242,
+      "ij": 243,
+      "Ĵ": 244,
+      "ĵ": 245,
+      "Ķ": 246,
+      "ķ": 247,
+      "ĸ": 248,
+      "Ĺ": 249,
+      "ĺ": 250,
+      "Ļ": 251,
+      "ļ": 252,
+      "Ľ": 253,
+      "ľ": 254,
+      "Ŀ": 255,
+      "ŀ": 256,
+      "Ł": 257,
+      "ł": 258,
+      "Ń": 259,
+      "he": 260,
+      "ll": 261,
+      "hell": 262,
+      "te": 263,
+      "<|unused264|>": 264,
+      "<|unused265|>": 265,
+      "<|unused266|>": 266,
+      "<|unused267|>": 267,
+      "<|unused268|>": 268,
+      "<|unused269|>": 269,
+      "<|unused270|>": 270,
+      "<|unused271|>": 271,
+      "<|unused272|>": 272,
+      "<|unused273|>": 273,
+      "<|unused274|>": 274,
+      "<|unused275|>": 275,
+      "<|unused276|>": 276,
+      "<|unused277|>": 277,
+      "<|unused278|>": 278,
+      "<|unused279|>": 279,
+      "<|unused280|>": 280,
+      "<|unused281|>": 281,
+      "<|unused282|>": 282,
+      "<|unused283|>": 283,
+      "<|unused284|>": 284,
+      "<|unused285|>": 285,
+      "<|unused286|>": 286,
+      "<|unused287|>": 287
+    },
+    "merges": [
+      [
+        "h",
+        "e"
+      ],
+      [
+        "l",
+        "l"
+      ],
+      [
+        "he",
+        "ll"
+      ],
+      [
+        "t",
+        "e"
+      ]
+    ]
+  }
+}
\ No newline at end of file
diff --git a/tests/fixtures/kv-shared-absent/tokenizer_config.json b/tests/fixtures/kv-shared-absent/tokenizer_config.json
new file mode 100644
index 00000000..ee42569a
--- /dev/null
+++ b/tests/fixtures/kv-shared-absent/tokenizer_config.json
@@ -0,0 +1,8 @@
+{
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|im_end|>",
+  "pad_token": "<pad>",
+  "unk_token": "<|endoftext|>",
+  "chat_template": "{% for m in messages %}<|im_start|>{{ m['role'] }}\n{{ m['content'] }}<|im_end|>\n{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"
+}
\ No newline at end of file
diff --git a/tests/fixtures/kv-shared-present/config.json b/tests/fixtures/kv-shared-present/config.json
new file mode 100644
index 00000000..78cfc80b
--- /dev/null
+++ b/tests/fixtures/kv-shared-present/config.json
@@ -0,0 +1,28 @@
+{
+  "model_type": "gemma4_text",
+  "architectures": [
+    "Gemma4ForCausalLM"
+  ],
+  "hidden_size": 64,
+  "num_hidden_layers": 4,
+  "intermediate_size": 128,
+  "num_attention_heads": 4,
+  "head_dim": 16,
+  "global_head_dim": 16,
+  "rms_norm_eps": 1e-06,
+  "vocab_size": 288,
+  "num_key_value_heads": 2,
+  "rope_traditional": false,
+  "rope_theta": 10000.0,
+  "sliding_window": 128,
+  "sliding_window_pattern": 1,
+  "max_position_embeddings": 512,
+  "num_kv_shared_layers": 2,
+  "use_double_wide_mlp": false,
+  "tie_word_embeddings": true,
+  "hidden_size_per_layer_input": 32,
+  "vocab_size_per_layer_input": 16,
+  "final_logit_softcapping": 30.0,
+  "enable_moe_block": false,
+  "attention_k_eq_v": false
+}
\ No newline at end of file
diff --git a/tests/fixtures/kv-shared-present/model.safetensors b/tests/fixtures/kv-shared-present/model.safetensors
new file mode 100644
index 00000000..ecab1a6e
Binary files /dev/null and b/tests/fixtures/kv-shared-present/model.safetensors differ
diff --git a/tests/fixtures/kv-shared-present/tokenizer.json b/tests/fixtures/kv-shared-present/tokenizer.json
new file mode 100644
index 00000000..21c0e134
--- /dev/null
+++ b/tests/fixtures/kv-shared-present/tokenizer.json
@@ -0,0 +1,380 @@
+{
+  "version": "1.0",
+  "truncation": null,
+  "padding": null,
+  "added_tokens": [
+    {
+      "id": 0,
+      "content": "<|endoftext|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 1,
+      "content": "<|im_start|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 2,
+      "content": "<|im_end|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 3,
+      "content": "<pad>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    }
+  ],
+  "normalizer": null,
+  "pre_tokenizer": {
+    "type": "ByteLevel",
+    "add_prefix_space": false,
+    "trim_offsets": true,
+    "use_regex": true
+  },
+  "post_processor": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": false,
+    "use_regex": true
+  },
+  "decoder": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": true,
+    "use_regex": true
+  },
+  "model": {
+    "type": "BPE",
+    "dropout": null,
+    "unk_token": null,
+    "continuing_subword_prefix": null,
+    "end_of_word_suffix": null,
+    "fuse_unk": false,
+    "byte_fallback": false,
+    "ignore_merges": false,
+    "vocab": {
+      "<|endoftext|>": 0,
+      "<|im_start|>": 1,
+      "<|im_end|>": 2,
+      "<pad>": 3,
+      "!": 4,
+      "\"": 5,
+      "#": 6,
+      "$": 7,
+      "%": 8,
+      "&": 9,
+      "'": 10,
+      "(": 11,
+      ")": 12,
+      "*": 13,
+      "+": 14,
+      ",": 15,
+      "-": 16,
+      ".": 17,
+      "/": 18,
+      "0": 19,
+      "1": 20,
+      "2": 21,
+      "3": 22,
+      "4": 23,
+      "5": 24,
+      "6": 25,
+      "7": 26,
+      "8": 27,
+      "9": 28,
+      ":": 29,
+      ";": 30,
+      "<": 31,
+      "=": 32,
+      ">": 33,
+      "?": 34,
+      "@": 35,
+      "A": 36,
+      "B": 37,
+      "C": 38,
+      "D": 39,
+      "E": 40,
+      "F": 41,
+      "G": 42,
+      "H": 43,
+      "I": 44,
+      "J": 45,
+      "K": 46,
+      "L": 47,
+      "M": 48,
+      "N": 49,
+      "O": 50,
+      "P": 51,
+      "Q": 52,
+      "R": 53,
+      "S": 54,
+      "T": 55,
+      "U": 56,
+      "V": 57,
+      "W": 58,
+      "X": 59,
+      "Y": 60,
+      "Z": 61,
+      "[": 62,
+      "\\": 63,
+      "]": 64,
+      "^": 65,
+      "_": 66,
+      "`": 67,
+      "a": 68,
+      "b": 69,
+      "c": 70,
+      "d": 71,
+      "e": 72,
+      "f": 73,
+      "g": 74,
+      "h": 75,
+      "i": 76,
+      "j": 77,
+      "k": 78,
+      "l": 79,
+      "m": 80,
+      "n": 81,
+      "o": 82,
+      "p": 83,
+      "q": 84,
+      "r": 85,
+      "s": 86,
+      "t": 87,
+      "u": 88,
+      "v": 89,
+      "w": 90,
+      "x": 91,
+      "y": 92,
+      "z": 93,
+      "{": 94,
+      "|": 95,
+      "}": 96,
+      "~": 97,
+      "¡": 98,
+      "¢": 99,
+      "£": 100,
+      "¤": 101,
+      "¥": 102,
+      "¦": 103,
+      "§": 104,
+      "¨": 105,
+      "©": 106,
+      "ª": 107,
+      "«": 108,
+      "¬": 109,
+      "®": 110,
+      "¯": 111,
+      "°": 112,
+      "±": 113,
+      "²": 114,
+      "³": 115,
+      "´": 116,
+      "µ": 117,
+      "¶": 118,
+      "·": 119,
+      "¸": 120,
+      "¹": 121,
+      "º": 122,
+      "»": 123,
+      "¼": 124,
+      "½": 125,
+      "¾": 126,
+      "¿": 127,
+      "À": 128,
+      "Á": 129,
+      "Â": 130,
+      "Ã": 131,
+      "Ä": 132,
+      "Å": 133,
+      "Æ": 134,
+      "Ç": 135,
+      "È": 136,
+      "É": 137,
+      "Ê": 138,
+      "Ë": 139,
+      "Ì": 140,
+      "Í": 141,
+      "Î": 142,
+      "Ï": 143,
+      "Ð": 144,
+      "Ñ": 145,
+      "Ò": 146,
+      "Ó": 147,
+      "Ô": 148,
+      "Õ": 149,
+      "Ö": 150,
+      "×": 151,
+      "Ø": 152,
+      "Ù": 153,
+      "Ú": 154,
+      "Û": 155,
+      "Ü": 156,
+      "Ý": 157,
+      "Þ": 158,
+      "ß": 159,
+      "à": 160,
+      "á": 161,
+      "â": 162,
+      "ã": 163,
+      "ä": 164,
+      "å": 165,
+      "æ": 166,
+      "ç": 167,
+      "è": 168,
+      "é": 169,
+      "ê": 170,
+      "ë": 171,
+      "ì": 172,
+      "í": 173,
+      "î": 174,
+      "ï": 175,
+      "ð": 176,
+      "ñ": 177,
+      "ò": 178,
+      "ó": 179,
+      "ô": 180,
+      "õ": 181,
+      "ö": 182,
+      "÷": 183,
+      "ø": 184,
+      "ù": 185,
+      "ú": 186,
+      "û": 187,
+      "ü": 188,
+      "ý": 189,
+      "þ": 190,
+      "ÿ": 191,
+      "Ā": 192,
+      "ā": 193,
+      "Ă": 194,
+      "ă": 195,
+      "Ą": 196,
+      "ą": 197,
+      "Ć": 198,
+      "ć": 199,
+      "Ĉ": 200,
+      "ĉ": 201,
+      "Ċ": 202,
+      "ċ": 203,
+      "Č": 204,
+      "č": 205,
+      "Ď": 206,
+      "ď": 207,
+      "Đ": 208,
+      "đ": 209,
+      "Ē": 210,
+      "ē": 211,
+      "Ĕ": 212,
+      "ĕ": 213,
+      "Ė": 214,
+      "ė": 215,
+      "Ę": 216,
+      "ę": 217,
+      "Ě": 218,
+      "ě": 219,
+      "Ĝ": 220,
+      "ĝ": 221,
+      "Ğ": 222,
+      "ğ": 223,
+      "Ġ": 224,
+      "ġ": 225,
+      "Ģ": 226,
+      "ģ": 227,
+      "Ĥ": 228,
+      "ĥ": 229,
+      "Ħ": 230,
+      "ħ": 231,
+      "Ĩ": 232,
+      "ĩ": 233,
+      "Ī": 234,
+      "ī": 235,
+      "Ĭ": 236,
+      "ĭ": 237,
+      "Į": 238,
+      "į": 239,
+      "İ": 240,
+      "ı": 241,
+      "IJ": 242,
+      "ij": 243,
+      "Ĵ": 244,
+      "ĵ": 245,
+      "Ķ": 246,
+      "ķ": 247,
+      "ĸ": 248,
+      "Ĺ": 249,
+      "ĺ": 250,
+      "Ļ": 251,
+      "ļ": 252,
+      "Ľ": 253,
+      "ľ": 254,
+      "Ŀ": 255,
+      "ŀ": 256,
+      "Ł": 257,
+      "ł": 258,
+      "Ń": 259,
+      "he": 260,
+      "ll": 261,
+      "hell": 262,
+      "te": 263,
+      "<|unused264|>": 264,
+      "<|unused265|>": 265,
+      "<|unused266|>": 266,
+      "<|unused267|>": 267,
+      "<|unused268|>": 268,
+      "<|unused269|>": 269,
+      "<|unused270|>": 270,
+      "<|unused271|>": 271,
+      "<|unused272|>": 272,
+      "<|unused273|>": 273,
+      "<|unused274|>": 274,
+      "<|unused275|>": 275,
+      "<|unused276|>": 276,
+      "<|unused277|>": 277,
+      "<|unused278|>": 278,
+      "<|unused279|>": 279,
+      "<|unused280|>": 280,
+      "<|unused281|>": 281,
+      "<|unused282|>": 282,
+      "<|unused283|>": 283,
+      "<|unused284|>": 284,
+      "<|unused285|>": 285,
+      "<|unused286|>": 286,
+      "<|unused287|>": 287
+    },
+    "merges": [
+      [
+        "h",
+        "e"
+      ],
+      [
+        "l",
+        "l"
+      ],
+      [
+        "he",
+        "ll"
+      ],
+      [
+        "t",
+        "e"
+      ]
+    ]
+  }
+}
\ No newline at end of file
diff --git a/tests/fixtures/kv-shared-present/tokenizer_config.json b/tests/fixtures/kv-shared-present/tokenizer_config.json
new file mode 100644
index 00000000..ee42569a
--- /dev/null
+++ b/tests/fixtures/kv-shared-present/tokenizer_config.json
@@ -0,0 +1,8 @@
+{
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|im_end|>",
+  "pad_token": "<pad>",
+  "unk_token": "<|endoftext|>",
+  "chat_template": "{% for m in messages %}<|im_start|>{{ m['role'] }}\n{{ m['content'] }}<|im_end|>\n{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"
+}
\ No newline at end of file
diff --git a/tests/fixtures/stray-shard/config.json b/tests/fixtures/stray-shard/config.json
new file mode 100644
index 00000000..5f3594c0
--- /dev/null
+++ b/tests/fixtures/stray-shard/config.json
@@ -0,0 +1,16 @@
+{
+  "model_type": "qwen2",
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "vocab_size": 288,
+  "hidden_size": 64,
+  "intermediate_size": 128,
+  "num_hidden_layers": 2,
+  "num_attention_heads": 4,
+  "num_key_value_heads": 2,
+  "max_position_embeddings": 512,
+  "rms_norm_eps": 1e-06,
+  "rope_theta": 10000.0,
+  "tie_word_embeddings": false
+}
\ No newline at end of file
diff --git a/tests/fixtures/stray-shard/extra-not-in-index.safetensors b/tests/fixtures/stray-shard/extra-not-in-index.safetensors
new file mode 100644
index 00000000..78bbfc4a
Binary files /dev/null and b/tests/fixtures/stray-shard/extra-not-in-index.safetensors differ
diff --git a/tests/fixtures/stray-shard/model.safetensors b/tests/fixtures/stray-shard/model.safetensors
new file mode 100644
index 00000000..6d98f758
Binary files /dev/null and b/tests/fixtures/stray-shard/model.safetensors differ
diff --git a/tests/fixtures/stray-shard/model.safetensors.index.json b/tests/fixtures/stray-shard/model.safetensors.index.json
new file mode 100644
index 00000000..b9ffbbec
--- /dev/null
+++ b/tests/fixtures/stray-shard/model.safetensors.index.json
@@ -0,0 +1 @@
+{"metadata": {"total_size": 0}, "weight_map": {"model.embed_tokens.weight": "model.safetensors", "model.norm.weight": "model.safetensors", "lm_head.weight": "model.safetensors", "model.layers.0.self_attn.q_proj.weight": "model.safetensors", "model.layers.0.self_attn.q_proj.bias": "model.safetensors", "model.layers.0.self_attn.k_proj.weight": "model.safetensors", "model.layers.0.self_attn.k_proj.bias": "model.safetensors", "model.layers.0.self_attn.v_proj.weight": "model.safetensors", "model.layers.0.self_attn.v_proj.bias": "model.safetensors", "model.layers.0.self_attn.o_proj.weight": "model.safetensors", "model.layers.0.mlp.gate_proj.weight": "model.safetensors", "model.layers.0.mlp.up_proj.weight": "model.safetensors", "model.layers.0.mlp.down_proj.weight": "model.safetensors", "model.layers.0.input_layernorm.weight": "model.safetensors", "model.layers.0.post_attention_layernorm.weight": "model.safetensors", "model.layers.1.self_attn.q_proj.weight": "model.safetensors", "model.layers.1.self_attn.q_proj.bias": "model.safetensors", "model.layers.1.self_attn.k_proj.weight": "model.safetensors", "model.layers.1.self_attn.k_proj.bias": "model.safetensors", "model.layers.1.self_attn.v_proj.weight": "model.safetensors", "model.layers.1.self_attn.v_proj.bias": "model.safetensors", "model.layers.1.self_attn.o_proj.weight": "model.safetensors", "model.layers.1.mlp.gate_proj.weight": "model.safetensors", "model.layers.1.mlp.up_proj.weight": "model.safetensors", "model.layers.1.mlp.down_proj.weight": "model.safetensors", "model.layers.1.input_layernorm.weight": "model.safetensors", "model.layers.1.post_attention_layernorm.weight": "model.safetensors"}}
\ No newline at end of file
diff --git a/tests/fixtures/stray-shard/tokenizer.json b/tests/fixtures/stray-shard/tokenizer.json
new file mode 100644
index 00000000..21c0e134
--- /dev/null
+++ b/tests/fixtures/stray-shard/tokenizer.json
@@ -0,0 +1,380 @@
+{
+  "version": "1.0",
+  "truncation": null,
+  "padding": null,
+  "added_tokens": [
+    {
+      "id": 0,
+      "content": "<|endoftext|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 1,
+      "content": "<|im_start|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 2,
+      "content": "<|im_end|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 3,
+      "content": "<pad>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    }
+  ],
+  "normalizer": null,
+  "pre_tokenizer": {
+    "type": "ByteLevel",
+    "add_prefix_space": false,
+    "trim_offsets": true,
+    "use_regex": true
+  },
+  "post_processor": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": false,
+    "use_regex": true
+  },
+  "decoder": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": true,
+    "use_regex": true
+  },
+  "model": {
+    "type": "BPE",
+    "dropout": null,
+    "unk_token": null,
+    "continuing_subword_prefix": null,
+    "end_of_word_suffix": null,
+    "fuse_unk": false,
+    "byte_fallback": false,
+    "ignore_merges": false,
+    "vocab": {
+      "<|endoftext|>": 0,
+      "<|im_start|>": 1,
+      "<|im_end|>": 2,
+      "<pad>": 3,
+      "!": 4,
+      "\"": 5,
+      "#": 6,
+      "$": 7,
+      "%": 8,
+      "&": 9,
+      "'": 10,
+      "(": 11,
+      ")": 12,
+      "*": 13,
+      "+": 14,
+      ",": 15,
+      "-": 16,
+      ".": 17,
+      "/": 18,
+      "0": 19,
+      "1": 20,
+      "2": 21,
+      "3": 22,
+      "4": 23,
+      "5": 24,
+      "6": 25,
+      "7": 26,
+      "8": 27,
+      "9": 28,
+      ":": 29,
+      ";": 30,
+      "<": 31,
+      "=": 32,
+      ">": 33,
+      "?": 34,
+      "@": 35,
+      "A": 36,
+      "B": 37,
+      "C": 38,
+      "D": 39,
+      "E": 40,
+      "F": 41,
+      "G": 42,
+      "H": 43,
+      "I": 44,
+      "J": 45,
+      "K": 46,
+      "L": 47,
+      "M": 48,
+      "N": 49,
+      "O": 50,
+      "P": 51,
+      "Q": 52,
+      "R": 53,
+      "S": 54,
+      "T": 55,
+      "U": 56,
+      "V": 57,
+      "W": 58,
+      "X": 59,
+      "Y": 60,
+      "Z": 61,
+      "[": 62,
+      "\\": 63,
+      "]": 64,
+      "^": 65,
+      "_": 66,
+      "`": 67,
+      "a": 68,
+      "b": 69,
+      "c": 70,
+      "d": 71,
+      "e": 72,
+      "f": 73,
+      "g": 74,
+      "h": 75,
+      "i": 76,
+      "j": 77,
+      "k": 78,
+      "l": 79,
+      "m": 80,
+      "n": 81,
+      "o": 82,
+      "p": 83,
+      "q": 84,
+      "r": 85,
+      "s": 86,
+      "t": 87,
+      "u": 88,
+      "v": 89,
+      "w": 90,
+      "x": 91,
+      "y": 92,
+      "z": 93,
+      "{": 94,
+      "|": 95,
+      "}": 96,
+      "~": 97,
+      "¡": 98,
+      "¢": 99,
+      "£": 100,
+      "¤": 101,
+      "¥": 102,
+      "¦": 103,
+      "§": 104,
+      "¨": 105,
+      "©": 106,
+      "ª": 107,
+      "«": 108,
+      "¬": 109,
+      "®": 110,
+      "¯": 111,
+      "°": 112,
+      "±": 113,
+      "²": 114,
+      "³": 115,
+      "´": 116,
+      "µ": 117,
+      "¶": 118,
+      "·": 119,
+      "¸": 120,
+      "¹": 121,
+      "º": 122,
+      "»": 123,
+      "¼": 124,
+      "½": 125,
+      "¾": 126,
+      "¿": 127,
+      "À": 128,
+      "Á": 129,
+      "Â": 130,
+      "Ã": 131,
+      "Ä": 132,
+      "Å": 133,
+      "Æ": 134,
+      "Ç": 135,
+      "È": 136,
+      "É": 137,
+      "Ê": 138,
+      "Ë": 139,
+      "Ì": 140,
+      "Í": 141,
+      "Î": 142,
+      "Ï": 143,
+      "Ð": 144,
+      "Ñ": 145,
+      "Ò": 146,
+      "Ó": 147,
+      "Ô": 148,
+      "Õ": 149,
+      "Ö": 150,
+      "×": 151,
+      "Ø": 152,
+      "Ù": 153,
+      "Ú": 154,
+      "Û": 155,
+      "Ü": 156,
+      "Ý": 157,
+      "Þ": 158,
+      "ß": 159,
+      "à": 160,
+      "á": 161,
+      "â": 162,
+      "ã": 163,
+      "ä": 164,
+      "å": 165,
+      "æ": 166,
+      "ç": 167,
+      "è": 168,
+      "é": 169,
+      "ê": 170,
+      "ë": 171,
+      "ì": 172,
+      "í": 173,
+      "î": 174,
+      "ï": 175,
+      "ð": 176,
+      "ñ": 177,
+      "ò": 178,
+      "ó": 179,
+      "ô": 180,
+      "õ": 181,
+      "ö": 182,
+      "÷": 183,
+      "ø": 184,
+      "ù": 185,
+      "ú": 186,
+      "û": 187,
+      "ü": 188,
+      "ý": 189,
+      "þ": 190,
+      "ÿ": 191,
+      "Ā": 192,
+      "ā": 193,
+      "Ă": 194,
+      "ă": 195,
+      "Ą": 196,
+      "ą": 197,
+      "Ć": 198,
+      "ć": 199,
+      "Ĉ": 200,
+      "ĉ": 201,
+      "Ċ": 202,
+      "ċ": 203,
+      "Č": 204,
+      "č": 205,
+      "Ď": 206,
+      "ď": 207,
+      "Đ": 208,
+      "đ": 209,
+      "Ē": 210,
+      "ē": 211,
+      "Ĕ": 212,
+      "ĕ": 213,
+      "Ė": 214,
+      "ė": 215,
+      "Ę": 216,
+      "ę": 217,
+      "Ě": 218,
+      "ě": 219,
+      "Ĝ": 220,
+      "ĝ": 221,
+      "Ğ": 222,
+      "ğ": 223,
+      "Ġ": 224,
+      "ġ": 225,
+      "Ģ": 226,
+      "ģ": 227,
+      "Ĥ": 228,
+      "ĥ": 229,
+      "Ħ": 230,
+      "ħ": 231,
+      "Ĩ": 232,
+      "ĩ": 233,
+      "Ī": 234,
+      "ī": 235,
+      "Ĭ": 236,
+      "ĭ": 237,
+      "Į": 238,
+      "į": 239,
+      "İ": 240,
+      "ı": 241,
+      "IJ": 242,
+      "ij": 243,
+      "Ĵ": 244,
+      "ĵ": 245,
+      "Ķ": 246,
+      "ķ": 247,
+      "ĸ": 248,
+      "Ĺ": 249,
+      "ĺ": 250,
+      "Ļ": 251,
+      "ļ": 252,
+      "Ľ": 253,
+      "ľ": 254,
+      "Ŀ": 255,
+      "ŀ": 256,
+      "Ł": 257,
+      "ł": 258,
+      "Ń": 259,
+      "he": 260,
+      "ll": 261,
+      "hell": 262,
+      "te": 263,
+      "<|unused264|>": 264,
+      "<|unused265|>": 265,
+      "<|unused266|>": 266,
+      "<|unused267|>": 267,
+      "<|unused268|>": 268,
+      "<|unused269|>": 269,
+      "<|unused270|>": 270,
+      "<|unused271|>": 271,
+      "<|unused272|>": 272,
+      "<|unused273|>": 273,
+      "<|unused274|>": 274,
+      "<|unused275|>": 275,
+      "<|unused276|>": 276,
+      "<|unused277|>": 277,
+      "<|unused278|>": 278,
+      "<|unused279|>": 279,
+      "<|unused280|>": 280,
+      "<|unused281|>": 281,
+      "<|unused282|>": 282,
+      "<|unused283|>": 283,
+      "<|unused284|>": 284,
+      "<|unused285|>": 285,
+      "<|unused286|>": 286,
+      "<|unused287|>": 287
+    },
+    "merges": [
+      [
+        "h",
+        "e"
+      ],
+      [
+        "l",
+        "l"
+      ],
+      [
+        "he",
+        "ll"
+      ],
+      [
+        "t",
+        "e"
+      ]
+    ]
+  }
+}
\ No newline at end of file
diff --git a/tests/fixtures/stray-shard/tokenizer_config.json b/tests/fixtures/stray-shard/tokenizer_config.json
new file mode 100644
index 00000000..ee42569a
--- /dev/null
+++ b/tests/fixtures/stray-shard/tokenizer_config.json
@@ -0,0 +1,8 @@
+{
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|im_end|>",
+  "pad_token": "<pad>",
+  "unk_token": "<|endoftext|>",
+  "chat_template": "{% for m in messages %}<|im_start|>{{ m['role'] }}\n{{ m['content'] }}<|im_end|>\n{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"
+}
\ No newline at end of file
diff --git a/tests/test-fixtures.sh b/tests/test-fixtures.sh
new file mode 100755
index 00000000..230fa3a0
--- /dev/null
+++ b/tests/test-fixtures.sh
@@ -0,0 +1,93 @@
+#!/usr/bin/env bash
+# test-fixtures.sh — load every synthetic checkpoint shape and generate a token.
+#
+# These fixtures are a few hundred kilobytes each and live in the repository, so this
+# needs no download, no model cache, and no network. See scripts/make-test-fixtures.py
+# for why they exist and what they do and do not cover.
+#
+# Each shape corresponds to a defect that reached users:
+#   dense               baseline — nothing special, catches gross breakage
+#   stray-shard         #118: a .safetensors beside the index but absent from it
+#   kv-shared-absent    #120: gemma-4-e4b shape, shared layers ship no k/v
+#   kv-shared-present   b674:  gemma-4-e2b shape, shared layers ship k/v anyway
+#
+# The output is gibberish by construction — the weights are random. A fixture passes
+# when the server loads it and produces *a* token, which is what exercises config
+# parsing, weight-key resolution, sanitisation and layer materialisation.
+#
+# Usage: ./tests/test-fixtures.sh [binary] [port]
+
+set -uo pipefail
+
+BINARY="${1:-.build/release/SwiftLM}"
+PORT="${2:-15460}"
+HOST="127.0.0.1"
+FIXTURE_DIR="$(cd "$(dirname "$0")" && pwd)/fixtures"
+
+GREEN='\033[0;32m'; RED='\033[0;31m'; YELLOW='\033[1;33m'; NC='\033[0m'
+PASS=0; FAIL=0
+
+log()  { echo -e "${YELLOW}[fixtures]${NC} $*"; }
+pass() { PASS=$((PASS + 1)); echo -e "  ${GREEN}✅ PASS${NC}: $*"; }
+fail() { FAIL=$((FAIL + 1)); echo -e "  ${RED}❌ FAIL${NC}: $*"; }
+
+SERVER_PID=""
+cleanup() { [ -n "$SERVER_PID" ] && kill "$SERVER_PID" 2>/dev/null; SERVER_PID=""; }
+trap cleanup EXIT
+
+run_fixture() {
+    local name="$1"
+    local dir="$FIXTURE_DIR/$name"
+    local url="http://$HOST:$PORT"
+    local logfile="/tmp/SwiftLM-test-fixture-$name.log"
+
+    if [ ! -d "$dir" ]; then
+        fail "$name: fixture directory missing — run scripts/make-test-fixtures.py"
+        return
+    fi
+
+    "$BINARY" --model "$dir" --port "$PORT" --host "$HOST" > "$logfile" 2>&1 &
+    SERVER_PID=$!
+
+    local ready=0
+    for _ in $(seq 1 60); do
+        if curl -sf "$url/health" >/dev/null 2>&1; then ready=1; break; fi
+        if ! kill -0 "$SERVER_PID" 2>/dev/null; then break; fi
+        sleep 1
+    done
+
+    if [ "$ready" -ne 1 ]; then
+        fail "$name: server did not start — $(grep -m1 -E '^Error|Fatal' "$logfile" || echo 'see '"$logfile")"
+        cleanup
+        return
+    fi
+
+    local body
+    body=$(curl -sf --max-time 60 "$url/v1/chat/completions" \
+        -H 'Content-Type: application/json' \
+        -d '{"messages":[{"role":"user","content":"ping"}],"max_tokens":4,"stream":false}' 2>/dev/null)
+
+    # Random weights make the text meaningless, so assert on the token count instead.
+    if [ -n "$body" ] && echo "$body" | python3 -c '
+import json, sys
+d = json.load(sys.stdin)
+sys.exit(0 if d["usage"]["completion_tokens"] >= 1 else 1)' 2>/dev/null; then
+        pass "$name loaded and generated"
+    else
+        fail "$name: no completion — $(echo "$body" | head -c 120)"
+    fi
+
+    cleanup
+    sleep 1
+}
+
+log "Binary: $BINARY"
+for name in dense stray-shard kv-shared-absent kv-shared-present; do
+    log "Shape: $name"
+    run_fixture "$name"
+done
+
+log "═══════════════════════════════════════"
+log "Results: $PASS passed, $FAIL failed"
+log "═══════════════════════════════════════"
+[ "$FAIL" -eq 0 ]
diff --git a/tests/test-vision.sh b/tests/test-vision.sh
index c33276d0..13e0f39a 100755
--- a/tests/test-vision.sh
+++ b/tests/test-vision.sh
@@ -89,8 +89,37 @@ mkdir -p /tmp/vision_test
 # 28x28 black PNG (requires multiple of 28 for Qwen2-VL patch embedder)
 BASE64_IMG="iVBORw0KGgoAAAANSUhEUgAAABwAAAAcCAIAAAD9b0jDAAAAGUlEQVR4nO3BMQEAAADCoPVPbQdvoAAA6DQJTAABRMAOLAAAAABJRU5ErkJggg=="
 
+# A model id names a moving target. On 2026-08-12 LiquidAI republished
+# LFM2.5-VL-450M-MLX-4bit with a chat template one brace short of valid —
+# `{- bos_token -}}` where the previous revision had `{{- bos_token -}}` — so every
+# request became `parser('Unexpected token type: closeExpression')`, HTTP 500, and main
+# went red five minutes later with nothing changed on our side. Restoring that single
+# brace locally makes the same revision answer normally, so the fault is upstream, not
+# a compatibility gap worth chasing here.
+#
+# Pinning is what stops a third party's afternoon from deciding whether this repository
+# has a green build. Re-point it deliberately, when someone means to test a newer
+# revision, and treat the failure that follows as a real result.
+LFM_REPO="LiquidAI/LFM2.5-VL-450M-MLX-4bit"
+LFM_REVISION="10ce3604e42cd595497c47aaf67b7890e1e2a3b4"
+
+# Resolve to the pinned snapshot on disk. Falling back to the bare repo id keeps a local
+# `./tests/test-vision.sh` working without a prefetch, but says so — a run that quietly
+# tested a different revision than CI did is worse than one that took a moment longer.
+pinned_snapshot() {
+    local repo="$1" revision="$2"
+    local hub="${HF_HUB_CACHE:-${HF_HOME:-$HOME/.cache/huggingface}/hub}"
+    local dir="$hub/models--${repo//\//--}/snapshots/$revision"
+    if [ -d "$dir" ]; then
+        echo "$dir"
+    else
+        log "note: pinned revision ${revision:0:7} of $repo is not in the cache; using the floating id"
+        echo "$repo"
+    fi
+}
+
 run_case "mlx-community/Qwen2-VL-2B-Instruct-4bit" "$BASE_PORT" "yes"
-run_case "LiquidAI/LFM2.5-VL-450M-MLX-4bit" "$((BASE_PORT + 1))" "yes"
+run_case "$(pinned_snapshot "$LFM_REPO" "$LFM_REVISION")" "$((BASE_PORT + 1))" "yes"
 
 rm -rf /tmp/vision_test
 exit 0