feat(latent): add learned upscaling and captured noise reuse

Add provenance recovery, coordinated native sampler-state capture, frequency-completed noise reuse, and optional full-resolution inversion finishing. Support automatic SesquiLSR selection and native bislerp fallback for unknown latents.

Verify exported Anima and Krea CD workflows with zero VAE encodes. Finishing outputs reproduce prior images exactly; reuse-only outputs match independent references at the agreed capture coordinates. All 3813 Python tests and 115 frontend tests pass alongside required quality gates.

tests/node_api/test_latent_reuse_nodes.py

@@ -0,0 +1,74 @@
+"""Verify the public latent reuse node contracts and v3 registration."""
+
+from typing import Any
+
+import pytest
+
+from simple_syrup.domain.sampler_options import (
+    ContextualDiffusionOptions,
+    SamplerOptions,
+)
+from simple_syrup.nodes_v3 import get_nodes
+from simple_syrup.nodes_v3.latent_from_image import LatentFromImageV3
+from simple_syrup.nodes_v3.noise_reuse_options import NoiseReuseOptionsV3
+from simple_syrup.nodes_v3.simple_upscale_latent import SimpleUpscaleLatentV3
+from simple_syrup.services.latent_provenance_service import LatentProvenanceService
+
+
+@pytest.mark.parametrize(
+    "node", (LatentFromImageV3, NoiseReuseOptionsV3, SimpleUpscaleLatentV3)
+)
+def test_v3_registration_and_tooltips_cover_public_contract(node: Any) -> None:
+    """Export all three nodes with descriptions and actionable socket tooltips."""
+    assert node in get_nodes()
+    schema = node.define_schema()
+    assert schema.node_id.startswith("SimpleSyrup.")
+    assert schema.description
+    assert all(value.tooltip for value in (*schema.inputs, *schema.outputs))
+
+
+def test_upscale_exposes_one_dropdown_and_scale_default() -> None:
+    """Keep format and model loading choices behind the agreed method control."""
+    schema = SimpleUpscaleLatentV3.define_schema()
+    inputs = {value.id: value for value in schema.inputs}
+    assert list(inputs) == ["latent", "method", "scale_factor"]
+    assert inputs["method"].default == "Auto"
+    assert "bislerp" in inputs["method"].tooltip.lower()
+    assert inputs["scale_factor"].default == 2.0
+    assert schema.display_name == "Simple Upscale Latent"
+
+
+def test_reuse_enabled_is_first_widget_and_disabled_passes_same_prefix() -> None:
+    """Prove zero finishing is reuse-only and switching off preserves prior options."""
+    schema = NoiseReuseOptionsV3.define_schema()
+    inputs = {value.id: value for value in schema.inputs}
+    assert list(inputs) == [
+        "options",
+        "enabled",
+        "finishing_steps",
+        "finishing_method",
+        "finishing_start_fraction",
+    ]
+    assert inputs["enabled"].default is True
+    assert inputs["finishing_steps"].default == 0
+    prefix = SamplerOptions(contextual_diffusion=ContextualDiffusionOptions())
+    assert NoiseReuseOptionsV3.execute(enabled=False, options=prefix)[0] is prefix
+    enabled = NoiseReuseOptionsV3.execute(options=prefix)[0]
+    assert enabled.contextual_diffusion is prefix.contextual_diffusion
+    assert enabled.noise_reuse is not None and enabled.noise_reuse.finishing_steps == 0
+
+
+def test_provenance_recovery_has_no_vae_or_widgets_and_never_encodes() -> None:
+    """Return graph lineage directly instead of producing a re-encoded approximation."""
+    schema = LatentFromImageV3.define_schema()
+    assert [value.id for value in schema.inputs] == ["image"]
+    service = LatentProvenanceService()
+    prompt = {
+        "decode": {
+            "class_type": "VAEDecode",
+            "inputs": {"samples": ["sample", 0], "vae": ["loader", 2]},
+        }
+    }
+    assert service.recover_link(prompt, ["decode", 0], {}) == ("sample", 0)
+    with pytest.raises(ValueError, match="unmodified"):
+        service.recover_link({}, ["loaded_image", 0], {})