v2.0.0: remove two no-op widgets, add streaming/interrupt, structured output, honest diagnostics

[EnragedAntelope]

Why

The lmstudio SDK silently discards prediction-config keys it does not recognise instead of raising. v1.x relied on that not happening, and shipped two widgets that never did anything:

Widget	Sent as	Reality
presence_penalty	presencePenalty	not in LlmPredictionConfigDict — dropped
enable_thinking	enableThinking	not in LlmPredictionConfigDict — dropped

Confirmed against a live LM Studio server: the prediction_config the server echoes back contains neither. LM Studio has no presence penalty, no frequency penalty, no inference-time seed and no thinking on/off flag — seed exists only in the load config.

To stop this recurring, generate() now diffs the config it sent against the config the server reports as applied, and warns about anything dropped.

Breaking change, mitigated

Both widgets are removed and the rest regrouped (sampling → output shaping → reasoning → vision → speculative decoding → management). ComfyUI serialises widgets_values positionally, so this would shift every later value — web/ea_lmstudio.js detects the v1.5.x layout and writes values onto current widgets by name, which is order-independent.

Browser-verified with a real v1 workflow: without the migration reasoning_mode loads as "[[/R]]" and custom_open_tag as false; with it, all 23 legacy values land correctly on the regrouped widgets and the four new ones take their defaults.

Fixes

• A successful generation could be reported as Generation failed. Every LlmPredictionStats field except stop_reason is Optional; formatting None with :.2f raised TypeError after the model had produced its text, which the outer handler turned into an error and discarded.
• Restored workflows drew widgets outside the node frame. A workflow stores the node's size and LiteGraph restores it verbatim without re-checking that the widgets fit. v2.0.0 has a net +2 widgets, so every upgraded workflow was too short. growToFitWidgets grows (never shrinks) the node — a v1 node stored at 700px now grows to 820px on load.
• New nodes defaulted to the -- Custom -- sentinel with an empty identifier, so the first run always failed even when discovery worked.
• Models whose identifier fails validation (LM Studio serves ids like model@?) vanished from the dropdown with no explanation.
• requires-python said >=3.9; the lmstudio dependency itself needs >=3.10.

Added

• Streamed prediction — ComfyUI's cancel button actually stops generation, and the queue progress bar tracks tokens.
• Native reasoning split — uses LM Studio's own reasoning_type fragment tagging when available; tag regexes remain the fallback.
• stop_strings, context_overflow, and structured JSON output (schema-constrained, with automatic ```json fence recovery).
• In-node response preview — OUTPUT_NODE now carries a ui payload. Excluded from both the API prompt and the saved workflow.
• Speculative-decoding acceptance stats plus a low-acceptance hint.
• example_workflows/ (discoverable in ComfyUI's template browser), AGENTS.md, and CI running the suite on 3.10 and 3.13.

Verification

• 102 tests pass (was 48).
• Live QA against LM Studio 0.4.x + SDK 1.5.0: text, VLM, stop strings, structured output, invalid schema, context policy, all sampling params, mid-stream cancel.
• Browser-verified: node registration, v1 migration onto regrouped widgets, in-node preview, serialization, widget/node overflow and node overlap in both example workflows, no console errors from this pack.

🤖 Generated with Claude Code