feat(renderers): add Inkling and Inkling-Small

[hallerite]

Summary

Adds one InklingRenderer shared by both Thinking Machines checkpoints:

• thinkingmachines/Inkling
• thinkingmachines/Inkling-Small

The two pinned checkpoints ship byte-identical chat templates, tokenizer assets, and processor configuration, so they intentionally use the same renderer and config. Each model ID is independently covered by the shared render, config, and multimodal parity matrices.

This PR is stacked on #123. Merge the Gemma 4 PR first, then this PR.

What it does

• Text, tools, reasoning, post-tool thinking, and reasoning_effort render with strict chat-template parity.
• Image and audio rendering matches the native InklingProcessor, including dynamic placeholder expansion and MultiModalData anchoring.
• parse_response preserves native JSON tool arguments and separates reasoning/content channels.
• bridge_to_next_turn carries multimodal state without mutating prior data.

Dependency

Keeps the repository-wide transformers>=4.50.0 floor so renderers remains co-installable with prime-rl's current Transformers pin. Inkling tokenizer and text-only rendering stay usable on older releases. Image/audio processing fails lazily with a clear upgrade message when native InklingProcessor support is unavailable (introduced in Transformers 5.14). The lock continues to resolve to transformers==5.14.1.

Review follow-up

• unknown content-part types and malformed/non-object tool arguments now raise instead of silently dropping data
• audio preprocessing uses the processor's public audio= call
• audio has an independent audio_cache_max bound
• both Inkling checkpoints now run in the shared bridge and round-trip matrices

Verification

• full combined Gemma 4 + Inkling stack: 3,331 passed, 173 skipped, 1 xfailed
• Ruff check, Ruff format check, and git diff --check pass
• focused Inkling bridge/round-trip suite: 59 passed, 2 expected skips
• Inkling multimodal suite: 13 passed, 4 expected skips
• all commits in the stack are signed with hallerite <git@hallerite.com> using the configured SSH signing key

---

Note

Medium Risk
Large new renderer (~1k lines) on the critical render/parse/bridge path; image/audio depend on Transformers 5.14 at runtime though the declared floor stays lower.

Overview
Adds InklingRenderer for thinkingmachines/Inkling and Inkling-Small, sharing one implementation because both checkpoints use identical template/tokenizer/processor assets.

The renderer matches apply_chat_template for text, tools, reasoning_effort (injected as Thinking effort level: {N}), and assistant reasoning blocks; image/audio paths expand placeholders like native InklingProcessor (Transformers ≥5.14) and attach MultiModalData. parse_inkling decodes completions into content, reasoning, and native-JSON tool calls. bridge_to_next_turn uses thinking retention all, merges multimodal sidecars without mutating prior data, and refuses extensions that need unresolved tool names.

InklingRendererConfig adds reasoning_effort, image_cache_max, and audio_cache_max. The package keeps transformers>=4.50.0 for co-install with prime-rl; multimodal fails lazily with an upgrade hint when InklingProcessor is missing. Docs, registry, and shared parity/bridge/multimodal tests are wired for both model IDs; uv.lock bumps to transformers==5.14.1.

^{Reviewed by Cursor Bugbot for commit de48898. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add Inkling and Inkling-Small renderers with image/audio multimodal support

• Introduces InklingRenderer in renderers/inkling.py with full rendering, parsing, bridging, and multimodal (image + audio) support for thinkingmachines/Inkling and thinkingmachines/Inkling-Small.
• Adds Gemma4Renderer in renderers/gemma4.py with image support and model-specific empty-thought prefill handling for google/gemma-4-* checkpoints.
• Registers both renderers in renderers/base.py and renderers/init.py with auto-resolution from model name and multimodal capability declarations.
• Adds InklingRendererConfig (with reasoning_effort, image_cache_max, audio_cache_max) and Gemma4RendererConfig (with enable_thinking, preserve_thinking, image_cache_max) to renderers/configs.py.
• Adds parse_inkling to renderers/parsing.py for decoding Inkling model outputs into structured ParsedResponse with tool call extraction.
• Extends renderers/client.py with _build_gemma4_features to serialize Gemma 4 image inputs into vLLM-compatible kwargs; raises RuntimeError if torch or required vLLM modules are absent.
• Risk: Inkling image/audio processing and Gemma 4 image preprocessing require transformers>=5.14; text-only rendering works on older releases.

^{Macroscope summarized e22d00f.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR introduces a substantial new renderer (~1000 lines of code) supporting Inkling/Inkling-Small models with multimodal capabilities (image and audio), new configuration options, and new parsing logic. The scope and complexity of this new feature warrants human review.

No code changes detected at de48898. Prior analysis still applies.

^{You can customize Macroscope's approvability policy. Learn more.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 7b2a55b. Configure here.}