diff --git a/.agents/skills/adding-a-new-model/SKILL.md b/.agents/skills/adding-a-new-model/SKILL.md
index 8a0ca0f66..9ec1c8887 100644
--- a/.agents/skills/adding-a-new-model/SKILL.md
+++ b/.agents/skills/adding-a-new-model/SKILL.md
@@ -35,6 +35,8 @@ Read these when you need deeper detail on a specific topic:
 - Find a small checkpoint on HuggingFace Hub for testing
 - Have the HuggingFace `transformers` source available to reference the
   PyTorch implementation
+- Pin the checkpoint revision before collecting configs, weights, processors,
+  parity data, or goldens
 
 ## Step-by-step
 
@@ -219,6 +221,12 @@ Then add the new module to `src/mobius/_configs/per_model/__init__.py` so its
 side-effect registration runs at import time. The dispatcher filters hooks by
 the declared model_type strings, so unrelated models never see your hook.
 
+Match upstream config semantics, not just field names: preserve transformation
+order (rounding/scaling), explicit `None`, alias precedence, zero-as-disabled
+sentinels, and wrapped/unwrapped composite configs. Add focused tests for each
+nontrivial transform. Test both trusted/custom config classes and raw pinned
+JSON, especially nested decoder fields such as head/layer counts.
+
 ### 7. Write tests
 
 See the **writing-tests** skill for full details.  At minimum:
@@ -281,6 +289,7 @@ see the [quality-checklist skill](../quality-checklist/SKILL.md).
 - [ ] ORT GenAI test added to `tests/ort_genai_test.py` (text-generation and VLM models)
 - [ ] CLI build works (`mobius build --model ...`)
 - [ ] Multi-dtype correctness verified (fp32, fp16, bf16)
+- [ ] Pinned revision reaches every Hub/processor/weight/golden call
 
 **Note:** Default optimizer passes (CSE, deduplicate initializers, identity
 elimination, remove unused nodes/opsets) are applied automatically.
@@ -426,6 +435,18 @@ config-driven features for the offending subclass.
 > embedding table off-by-one), read
 > [`references/weight-preprocessing.md`](references/weight-preprocessing.md).
 
+### 8. Compatibility and optional dependency traps
+
+- Guard imports of symbols from unpinned dependencies when older supported
+  versions may not define them.
+- If remote code imports an optional package, first prove the selected model
+  path needs it. Scope any test-only shim to that unused import; do not add a
+  production fallback or require an irrelevant package.
+- When extending a public component, append optional parameters after existing
+  positional parameters and add a positional-call compatibility test.
+- Helpers with rank/shape contracts must receive a value of that rank; do not
+  rely on incidental compatibility from a higher-rank tensor.
+
 ## Reference examples
 
 | Complexity | File | Why |
diff --git a/.agents/skills/debugging-multimodal/SKILL.md b/.agents/skills/debugging-multimodal/SKILL.md
index 7977f661b..0e4df41ee 100644
--- a/.agents/skills/debugging-multimodal/SKILL.md
+++ b/.agents/skills/debugging-multimodal/SKILL.md
@@ -261,11 +261,11 @@ via `MOBIUS_ORT_LOWER_OPSET_FOR_EP=1`).  See `src/mobius/_flags.py`.
 
 ### Encoder input dtype alignment
 
-Encoder task inputs should be declared with `dtype=config.dtype` so
-entry tensors match the model compute dtype (float32/float16/bfloat16).
-In the current codebase, multimodal encoder task builders set encoder
-inputs directly to `config.dtype` (there is no `_cast_encoder_input()`
-helper in `src/mobius/tasks/_base.py`).
+Real vision/audio processors emit float32. Encoder graph inputs should
+therefore be float32 even for fp16/bf16 exports, with one Cast to
+`config.dtype` at graph entry. Verify this with an actual processor batch
+and a graph I/O dtype assertion; synthetic feeds can hide the mismatch.
+Some older task builders still use `config.dtype` and are not safe templates.
 
 ### GQA for KV-shared layers
 
diff --git a/.agents/skills/diffusion-models/SKILL.md b/.agents/skills/diffusion-models/SKILL.md
index 6a03d10ce..959de5499 100644
--- a/.agents/skills/diffusion-models/SKILL.md
+++ b/.agents/skills/diffusion-models/SKILL.md
@@ -353,6 +353,26 @@ Compare `named_parameters()` output with HF weight names. Use the
 techniques from the **weight-name-alignment** skill to minimize
 `preprocess_weights`.
 
+### 7. Prove pipeline executability
+
+A component graph or metadata file is not a runnable diffusion pipeline.
+Before advertising runtime support, execute the complete path:
+
+`image encoder/VAE -> latent packing -> denoiser loop + scheduler -> target token unpacking -> VAE decoder`
+
+Verify sample/output ranks, source-vs-target token slicing, CFG semantics, and
+the runtime's actual scheduler identifiers/equations. Reload final metadata and
+verify every file path, port, preprocessing step, loop edge, and postprocess.
+If the runtime cannot express the dataflow, reject it before emitting artifacts;
+keep direct component export separate.
+
+Validate required VAE statistics against latent channel count before graph
+construction (`len(mean) == len(std) == z_dim`, positive standard deviations).
+Keep model-specific component/task overrides data-driven rather than adding
+pipeline-name conditionals. Fold initializers after weight loading and mark
+denoiser/VAE roles by semantics, not names. Tests must use `os.path.join`/`Path`
+for temporary ONNX and profiling paths so cleanup works on POSIX and Windows.
+
 ## Naming conventions for diffusers models
 
 Diffusers uses different conventions than transformers:
diff --git a/.agents/skills/multi-agent-coordination/SKILL.md b/.agents/skills/multi-agent-coordination/SKILL.md
index eb13d3853..48d095de2 100644
--- a/.agents/skills/multi-agent-coordination/SKILL.md
+++ b/.agents/skills/multi-agent-coordination/SKILL.md
@@ -40,7 +40,8 @@ git fetch origin && git checkout <branch> && git pull
 
 ## 2. Commit Coordination Protocol
 
-Multiple agents committing to the same branch requires discipline to avoid divergence.
+Choose the branch strategy in §9 first. When multiple agents contribute to one
+branch, use this protocol to avoid divergence.
 
 **Rules**:
 1. **Pull before commit**: Always `git pull origin <branch> --rebase` immediately before committing.
@@ -67,11 +68,14 @@ Checks at each stage prevent problems from compounding across agents.
 - Verify the branch exists and is up to date.
 - Run the relevant tests to establish a baseline — know what was already failing before you touched anything.
 - Check for uncommitted changes left by previous agents in your worktree.
+- Verify `inspect.getfile(mobius)` points into the agent's worktree; editable
+  installs from another worktree invalidate all test evidence.
 
 ### After each commit
 - Run the affected tests immediately (don't batch; catch regressions early).
 - Verify the commit landed on the correct branch (`git log --oneline -3`).
 - Confirm the push succeeded.
+- Confirm the remote ref/PR head SHA, not just local `HEAD`.
 
 ### Before final review (lead audit)
 - Check ALL worktrees for unpushed commits or uncommitted changes.
@@ -85,6 +89,15 @@ Checks at each stage prevent problems from compounding across agents.
 - Final test run passes.
 - PR description updated with an accurate scorecard of what changed.
 
+### CI failures and main refreshes
+- Compare the exact failing test on the exact base SHA; a red overall workflow
+  does not prove the same job failed on base.
+- Keep feature scope: document unrelated baseline names/metrics instead of
+  changing their code or tolerances.
+- Rebase independent PRs linearly onto `origin/main`. Resolve shared registries,
+  exports, helpers, and guards semantically by retaining both sides; then run
+  shared-surface tests, lint, and review before `push --force-with-lease`.
+
 ---
 
 ## 4. Task Dependency Management (DAG)
@@ -171,18 +184,14 @@ Match the agent role to the task type. Mismatched roles waste agent capacity.
 
 ---
 
-## 9. Single-Branch Strategy
-
-For large multi-agent efforts, use **one shared feature branch** rather than one branch per agent.
-
-**Why**: Avoids complex multi-branch merge scenarios at the end. All agents commit to the same branch via their isolated worktrees.
-
-**Trade-offs**:
-- More `pull --rebase` cycles per agent.
-- Occasional push conflicts (recoverable with rebase).
-- BUT: much simpler final state, single PR, linear history.
+## 9. Branch Strategy
 
-**Alternative**: Separate branches per agent merged via separate PRs — only viable when features are truly independent and don't share files.
+- Use one branch only when agents contribute to one atomic PR; serialize edits
+  to shared files and require pull/rebase before every push.
+- Use separate branches/PRs for independent deliverables (for example, separate
+  model architectures). This isolates CI, review, rollback, and publication.
+- Never merge main into a linear-history feature branch. Rebase, resolve shared
+  surfaces semantically, and force-push only with `--force-with-lease`.
 
 ---
 
diff --git a/.agents/skills/multimodal-models/SKILL.md b/.agents/skills/multimodal-models/SKILL.md
index a69e698d7..156dff435 100644
--- a/.agents/skills/multimodal-models/SKILL.md
+++ b/.agents/skills/multimodal-models/SKILL.md
@@ -97,6 +97,16 @@ convention), not `vision_encoder`/`audio_encoder`.
 The embedding model must handle `num_image_tokens=0` (text-only input) by
 zero-padding `image_features` before Gather so indices stay in-bounds.
 
+Input mixing must also:
+- compute feature offsets over the flattened batch, not restart per row;
+- preserve separate image/video streams or define and test an explicit packed
+  order; and
+- cover two-row mixed-media prompts with distinguishable feature sentinels.
+
+For cross-attentive decoders, export total-length padding masks and only caches
+that are actually reusable. Do not apply a generic seq2seq cross-cache contract
+to fixed encoder states without proving projection/cache semantics.
+
 ### Conditional 3-or-4-model task
 
 Some models come in two tiers (e.g. Gemma4): small variants include an
@@ -219,72 +229,35 @@ models this overflows ORT's CUDA Gather kernel. **Workaround:** Split into
 L separate `Embedding([V, D])` tables via `nn.ModuleList`, and use `Slice`
 instead of `Gather` for per-layer projection indexing.
 
-## Vision/audio encoder f32 input casting
-
-> **This applies to ALL multimodal models and ALL inference paths** —
-> not just ORT GenAI, and not architecture-specific.
+## Processor-to-graph contract
 
-Image and audio preprocessing universally produces **float32** output.
-This is true across all frameworks and runtimes:
+Run the real processor before finalizing graph I/O. Record names, shapes,
+dtypes, media-row ordering, sampled frame positions, and empty-media behavior;
+synthetic packed video layouts often differ from processor-native rows.
 
-- **PIL / torchvision:** Pixel normalization outputs f32
-- **torchaudio / librosa:** Mel spectrograms are f32
-- **ORT GenAI image_processor:** Resize, normalize, tile → f32
-- **ORT GenAI audio_processor:** Feature extraction → f32
-- **ORT Python API:** Custom preprocessing pipelines → typically f32
-- **Foundry Local:** Uses GenAI processors → f32
-
-This means vision and audio encoder ONNX graphs must accept f32 inputs
-even when the model is built in f16 or bf16. The encoder adds a
-`Cast(f32 → model_dtype)` at its graph entry point so that any runtime
-can feed it preprocessed data without worrying about the model's
-internal precision.
-
-### How it works
+Vision/audio preprocessors emit float32. Encoder graphs therefore accept
+float32 at the boundary and cast once to fp16/bf16 internally:
 
 ```
-Input (f32 from ANY preprocessor — PIL, torchaudio, GenAI, etc.)
-    ↓
-Cast(to=FLOAT16)    ← inserted automatically by mobius
-    ↓
-Vision/Audio encoder (weights in f16/bf16)
-    ↓
-Output (model_dtype)
+processor f32 -> Cast(model dtype) -> reduced-precision encoder
 ```
 
-Encoder weights still use the requested dtype (f16/bf16) for memory
-efficiency — only the graph inputs are f32. The Cast is a lightweight
-op with negligible overhead.
-
-### Why f32 is the universal preprocessing dtype
+Test image-only, video-only, mixed media, two-row batches, and decode with zero
+new media. Build kwargs conditionally; strict processors may reject video-only
+arguments on image-only calls.
 
-Preprocessing involves floating-point arithmetic (mean subtraction,
-std division, resampling interpolation) where f32 is the natural
-precision. Converting to f16/bf16 before these operations would lose
-precision in the preprocessing itself. The model's internal precision
-only matters after the preprocessed data enters the encoder.
-
-### What mobius does
-
-Mobius always builds encoder graphs with f32 inputs — this is the
-default behavior, not gated behind any flag. It works correctly
-regardless of the inference runtime:
-
-- `--runtime ort-genai` → f32 inputs (GenAI processors output f32)
-- No `--runtime` flag → f32 inputs (ORT Python API, custom runtimes)
-- Foundry Local → f32 inputs (uses GenAI internally)
-
-Without the Cast-at-input, ORT throws a type mismatch error:
-```
-Type Error: Type parameter (T) bound to different types
-(tensor(float) and tensor(float16))
-```
+For composite checkpoints, generate processor assets from the parent config
+even if model construction exposes an unwrapped text sub-config; test both
+wrapped and unwrapped resolver forms.
 
-### For model authors
+For variable-length media, audit symbolic complexity: packed-attention metadata
+must scale with rows/windows, not total-patches × media-count. Keep dense masks
+only as a portable fallback and assert optimized CUDA/DML graphs lack the
+quadratic construction.
 
-If you're adding a new multimodal model, you don't need to handle this
-manually — mobius inserts the Cast automatically for all encoder graphs.
-If the model dtype is already f32, no Cast is needed.
+Config generation is not runtime support. Execute real media through ORT GenAI;
+if it cannot supply a required encoder input or position-ID rank, reject export
+before writing artifacts and use the same evidence for any Foundry waiver.
 
 ## GQA for KV-shared layers (Gemma4)
 
diff --git a/.agents/skills/onnx-export-quantization/SKILL.md b/.agents/skills/onnx-export-quantization/SKILL.md
index 50617949d..72a8eb48f 100644
--- a/.agents/skills/onnx-export-quantization/SKILL.md
+++ b/.agents/skills/onnx-export-quantization/SKILL.md
@@ -322,6 +322,13 @@ FP32 when the model is BF16. Add `op.CastLike(result, input)` to
 ensure dtype consistency. See the `reusable-components` skill's
 section on precision behaviour.
 
+### 6. Attention rewrite schema mismatch
+
+Quantizers may rewrite Attention to a contrib op with fewer outputs or a
+different cache contract. Restrict the quantized op set when schemas do not
+match, then load and execute the result; successful conversion alone is not
+evidence. Document the narrowed recipe and unsupported rewrite explicitly.
+
 ## Testing quantized models
 
 ### L4: Golden data generation
diff --git a/.agents/skills/ort-genai-config/SKILL.md b/.agents/skills/ort-genai-config/SKILL.md
index f12959a82..cb94def19 100644
--- a/.agents/skills/ort-genai-config/SKILL.md
+++ b/.agents/skills/ort-genai-config/SKILL.md
@@ -244,6 +244,15 @@ decoder_inputs = {name: name for name in decoder_input_names}
 This means the genai config automatically adapts when `RemoveDeadGraphInputsPass`
 removes unused inputs (e.g. `position_ids` absorbed by GQA fusion).
 
+Hybrid cache metadata must preserve global layer indices across KV, convolution,
+and recurrent states. Derive slot count from the maximum
+`past_key_values.<index>.*` input index plus one; counting only `.key` inputs
+silently drops non-KV or sparsely indexed layers.
+
+CUDA Graph capture belongs on stable-shape autoregressive decoder sessions,
+not one-shot variable-shape vision/embedding stages. Keep an explicit decoder
+opt-out and validate capture together with shared KV buffers on the real model.
+
 > For the full generation flow, input routing, QwenImageProcessor output
 > tensors, and the multimodal processor factory, see
 > [`references/multimodal-pipeline.md`](references/multimodal-pipeline.md).
@@ -273,6 +282,13 @@ processor doesn't provide. Either:
 1. Compute them externally and inject via NamedTensors, or
 2. Modify the vision model to compute them from `image_grid_thw` internally
 
+### Config writes successfully but runtime cannot execute
+
+Config/schema success is not runtime support. Run load plus generation through
+the exported contract. If the runtime cannot route required feature inputs,
+position-ID rank, cache state, scheduler, or multimodal metadata, reject export
+before writing artifacts and report the exact runtime version/limitation.
+
 ### "input_ids size exceeds max length"
 
 For image prompts, the tokenized input_ids (including image_pad tokens) can
diff --git a/.agents/skills/quality-checklist/SKILL.md b/.agents/skills/quality-checklist/SKILL.md
index 776a9b691..9f5b6453e 100644
--- a/.agents/skills/quality-checklist/SKILL.md
+++ b/.agents/skills/quality-checklist/SKILL.md
@@ -32,15 +32,16 @@ before the PR is merged.
       the model is not a standard text-generation model
 - [ ] `preprocess_weights()` correctly maps every HuggingFace state-dict key
       to the ONNX initializer name (verified by the weight-alignment test)
-- [ ] All new components use `from mobius.components import ...` (public API),
-      not private submodule paths
+- [ ] Models/tasks import components from the public API; component submodules
+      import sibling primitives directly to avoid package-init cycles
 - [ ] No explicit protobuf operations anywhere in new code
       (`onnx.helper`, `onnx.TensorProto`, etc. are forbidden)
 - [ ] Tensor shapes are annotated in comments after non-trivial operations
 - [ ] Automated code review (Copilot/PR review) has been run and all
       findings are resolved or explicitly dismissed with a reason
-- [ ] `lintrunner -a` is clean — zero lint errors before merging
-      (`lintrunner f --output oneline --all-files` to auto-fix, then re-run to confirm)
+- [ ] Run the CI formatter first (`lintrunner f --output oneline --all-files`),
+      then `lintrunner -a`; confirm the pinned tools actually ran (an
+      uninitialized/no-op lintrunner is not evidence)
 
 ### 2. L1 — Graph builds
 
@@ -56,6 +57,8 @@ before the PR is merged.
 
 - [ ] YAML test case created at `testdata/cases/<category>/<model>.yaml`
 - [ ] `test_model_id` field set to a real HuggingFace model ID
+- [ ] Revision/trust flags are pinned and forwarded through config, processor,
+      weights, golden generation, and CLI build
 - [ ] Schema validates: `python -m pytest tests/yaml_schema_test.py`
 
 ### 4. L3 — Synthetic parity
@@ -87,6 +90,8 @@ before the PR is merged.
 - [ ] Generation golden file committed to
       `testdata/golden/<cat>/<model>_generation.json`
 - [ ] `python -m pytest tests/e2e_golden_test.py -m generation -k "<model>"` passes
+- [ ] Sequence length is asserted before exact token/frame comparison (no
+      prefix-only `zip` comparison)
 
 > **Speech-language models:** The golden generation script supports the
 > `speech-language` task type for models that process audio inputs (e.g.,
@@ -104,10 +109,13 @@ before the PR is merged.
 
 ### 7. Multi-dtype correctness
 
-- [ ] fp32 tests pass (target: exact token match in greedy generation)
-- [ ] fp16 tests pass (target: logit parity `atol=1e-2`; token match for
-      first N tokens)
-- [ ] bf16 tests pass (target: logit parity `atol=1e-2`)
+- [ ] Every claimed dtype runs real, nonzero input and checks semantic output,
+      not merely export/session creation
+- [ ] fp32 passes full-logit parity and deterministic semantic output
+- [ ] fp16/bf16 pass full-logit parity (`atol=1e-2`) plus
+      architecture-appropriate tokens, frame IDs, transcript, or image output
+- [ ] A dtype that executes inaccurately is rejected explicitly and is not the
+      default; never silently downgrade
 
 Use the example `--compare-hf --dtype f16/bf16` flag if a comparison script
 exists:
@@ -146,15 +154,23 @@ python examples/<model>_text_generation.py --compare-hf --dtype bf16
       uniform KV cache shapes; if so, explicitly override the generated
       `genai_config.json` before runtime validation rather than assuming
       the default generated setting is correct
-- [ ] Encoder inputs (vision/audio) are declared with `dtype=config.dtype`
-      at sub-model entry (no stale float32-only cast guidance)
+- [ ] Vision/audio graph inputs match the real processor (normally float32);
+      reduced-precision encoders cast once at graph entry
+- [ ] Representative graph evidence covers raw/post-Mobius/post-weight and
+      ORT EP-optimized op histograms; verify critical fusions, unexpected
+      Transpose nodes, and activation-vs-scalar metadata Memcpy nodes
+- [ ] Final fusion claims use loaded weights/constants; no-weight graphs cannot
+      prove initializer folding
 
 ### 9. ORT GenAI runtime
 
 - [ ] Model can be loaded with `ort_genai.Model(output_dir)` without error
-- [ ] Greedy text generation produces non-empty, coherent output
+- [ ] Actual generation with required media/features produces coherent output;
+      schema/config emission alone is not runtime support
 - [ ] ORT GenAI test added to `tests/ort_genai_test.py`
       (or confirmed covered by an existing parametrized test)
+- [ ] Structurally unsupported contracts fail before artifacts are emitted and
+      have a version-specific, evidence-based waiver
 
 Run the ORT GenAI integration test:
 
@@ -176,6 +192,8 @@ python -m pytest tests/ort_genai_test.py -m integration_slow -k "<model>" -sv
 - [ ] Model can be loaded from the exported ONNX package by Olive
 - [ ] INT4 / INT8 quantization runs to completion without errors
 - [ ] Quantized model produces non-degenerate output (coherent text)
+- [ ] Quantization uses only required execution providers if unrelated provider
+      registration fails, and evidence includes size, load, and inference
 - [ ] If quantization changes the graph structure (e.g. MatMulNBits), verify
       the `genai_config.json` still loads correctly in ORT GenAI
 
@@ -199,6 +217,21 @@ novel weight layouts (e.g. fused QKV, non-standard expert routing).
       a comment in the source file or the skill notes explains it
 - [ ] README model table updated if this is a significant new addition
 
+### 13. Publication and CI triage
+
+- [ ] Run specialist review after L1-L3, before expensive L4/L5, and again
+      after optimization or rebase changes
+- [ ] Rebase linearly onto `origin/main`; resolve shared registries/helpers
+      semantically, run shared-surface tests, and push with `--force-with-lease`
+- [ ] After rebase, rerun config generation and actual runtime load/generation;
+      parity-only tests do not catch changed cache/runtime contracts
+- [ ] Confirm the remote PR head SHA, mergeability, replacement lint, and
+      architecture checks after the final push
+- [ ] Triage red CI at check/job/test granularity against the exact base SHA;
+      record test names and metrics, and never change unrelated models
+- [ ] For expensive GPU goldens, link targeted L4 and L5 jobs; aggregate
+      timeout or runner failure is not model evidence
+
 ---
 
 ## Waiver policy
diff --git a/.agents/skills/reusable-components/SKILL.md b/.agents/skills/reusable-components/SKILL.md
index ac1d4c11e..3651d5263 100644
--- a/.agents/skills/reusable-components/SKILL.md
+++ b/.agents/skills/reusable-components/SKILL.md
@@ -51,6 +51,10 @@ an underscore prefix for local use:
 from mobius.components import Conv2d as _Conv2d, SiLU as _SiLU
 ```
 
+Inside `src/mobius/components/_*.py`, import sibling primitives directly
+(`mobius.components._common`, etc.) so a partially initialized public package
+cannot create a circular import.
+
 Model-specific compound blocks (e.g. `_TimestepEmbedding`, `_DiTBlock`,
 `_ResNetBlock2D`) remain in the model files they belong to.
 
@@ -210,6 +214,17 @@ position IDs instead of an explicit mask.
    `op.Cast(to=ir.DataType.FLOAT)`, compute, then cast back with `op.CastLike(result, input)`.
    For dtype-adaptive parameters, use `op.CastLike(param, reference)`.
 
+8. **Prefer canonical ONNX operators.** Emit standard `BatchNormalization`,
+   `RMSNormalization`, and activation ops when semantics match so ORT can fold
+   and fuse them. Scale-free RMSNorm still needs a schema-valid 1-D scale and
+   must preserve HF's fp32 variance semantics before casting back. Prefer
+   `stash_type=FLOAT` over a decomposed graph when it matches HF, and verify
+   normalize-in-fp32 → cast activation → apply gamma ordering. Keep a manual
+   form when preventing an incorrect provider fusion is intentional and tested.
+
+9. **Preserve public call compatibility.** Append new optional arguments after
+   existing positional parameters and add a positional-call regression test.
+
 ## ONNX op patterns overview
 
 Key patterns for building components:
diff --git a/.agents/skills/writing-tests/SKILL.md b/.agents/skills/writing-tests/SKILL.md
index 79da765c3..2f92dd4f0 100644
--- a/.agents/skills/writing-tests/SKILL.md
+++ b/.agents/skills/writing-tests/SKILL.md
@@ -239,6 +239,33 @@ integration test alongside any new custom function or Scan op.**
 - **Vision with real pixel values** — zeros don't exercise the encoder
 - **All dtypes** (f32, f16, bf16) — each can expose different bugs
 - **GPU when available** — different kernels on CUDA
+- **Actual package wiring** — feed each ONNX stage from the preceding ONNX
+  stage, not an HF intermediate that bypasses the integration under test
+- **Defaults and masks** — assert constructor/config defaults in emitted ONNX
+  attributes and test padding invariance across prefill plus cached decode
+- **Real processor contract** — record input names, shapes, dtypes, media-row
+  ordering, and sampled frame positions from nonzero image/video/audio data
+- **Batch and decode edges** — use two rows with distinguishable media
+  features, mixed modality order, and a decode step with zero new media
+
+### Golden tests must be reproducible and exact
+
+- Pin one revision through config, processor, weight shards, reference
+  generation, and ONNX build; test that plumbing forwards it.
+- Assert sequence lengths before exact token/frame comparison. For CTC, compare
+  the full argmax frame sequence and collapsed transcript.
+- Test image-only, video-only, and mixed media; pass processor kwargs only for
+  media that are present.
+- On hosted runners, isolate and eagerly delete each test's Hub, assets, and
+  Xet caches. Patching `HF_HOME` after `huggingface_hub` import is insufficient;
+  patch its imported cache constants too.
+- Run the exact L2 discovery path: YAML schema, `test_model_id`, revision, and
+  trust flags can be missed by model-local tests.
+- Run affected-model detection on the full diff before GPU CI. A new model
+  should select its targeted L4/L5 cases; distinguish all-model timeout or
+  runner termination from a model assertion failure.
+- Reference goldens must come from an independently invoked upstream pipeline,
+  never from the implementation under test or ad-hoc intermediate features.
 
 ### Recurrent state ≠ KV cache