feat(renderers): add Inkling renderer (text + image + audio)

[hallerite]

Summary

Adds InklingRenderer for thinkingmachines/Inkling — a vision+audio MoE whose chat template is unlike anything else in the repo: a flat, token-delimited stream (<|message_{role}|>…<|content_{kind}|>…<|end_message|>, no BOS), a one-off Thinking effort level: {N} system line emitted before the first non-system message, and <|content_model_end_sampling|> as the assistant close / eos.

Full multimodal (text + image + audio), byte-parity-verified.

What it does

• Text / tools / reasoning / reasoning_effort — byte-parity with apply_chat_template across all shapes: effort labels and raw floats, tool declarations, multi-tool-call turns, content=None, and tool-name resolution via tool_call_id.
• Image + audio — byte-parity with the native InklingProcessor, including its Qwen-VL-style placeholder expansion (image → num_patches × <|unused_200054|>, audio → one <|unused_200053|> per mel frame). Ships pixel_values / audio_input_ids in MultiModalData with exact placeholder anchoring.
• parse_response — native-JSON tool args (types preserved, no schema coercion needed), reasoning + content channels, tool-call token_spans.
• bridge_to_next_turn — carries MultiModalData forward; refuses tool messages whose name would need cross-turn resolution (would diverge from a full re-render).

Dependency change (please review)

The InklingProcessor / image / audio processors are native in transformers ≥ 5.14 (merged upstream 2026-07-15), so they load with no trust_remote_code. This bumps the shared transformers floor 5.6.2 → 5.14.1 (~8 minor versions). I re-ran the entire existing suite on 5.14.1 first — no regressions to the other 20+ renderers.

Because 5.14.x is inside the repo's 7-day exclude-newer cooldown, transformers is fast-tracked in [tool.uv.exclude-newer-package] so uv lock can resolve it today. This weakens the supply-chain cooldown for transformers — there's a comment noting the fast-track should be dropped once 5.14.x clears the window. Lock churn is minimal (only transformers + safetensors).

Tests

• New tests/test_inkling.py (31 tests): effort formatting, assistant structure/masks, tool-call shape, tool-name resolution, parse round-trips, bridge, config validation.
• Inkling added to the shared conftest render-parity barrage, test_renderer_config_parity (exercises reasoning_effort), and test_multimodal (new inkling family + audio modality kit).
• Full suite: 2710 passed, 114 skipped, 1 xfailed (2584 pre-Inkling). Multimodal parity across all families: 91 passed. ruff clean.

🤖 Generated with Claude Code

---

Note

Medium Risk
Large new renderer surface is well-tested, but raising the global transformers minimum (~8 minor versions) is a breaking install constraint for downstream stacks pinned to older versions.

Overview
Adds InklingRenderer for thinkingmachines/Inkling: token-delimited chat rendering (roles, thinking, tools, Thinking effort level system line), image and audio via native InklingProcessor placeholder expansion, MultiModalData sidecars, parse_inkling for completions, and bridge_to_next_turn with multimodal carry-forward (returns None when tool names need prior-turn resolution).

Wires InklingRendererConfig (reasoning_effort, image_cache_max), registry/model map, and documents Inkling as multimodal in README and docs/renderer-config.md (bridge policy all).

Dependency: bumps transformers>=5.14.0 (was ≥4.50/5.6) and fast-tracks transformers in uv exclude-newer-package until it clears the 7-day cooldown; lock updates include transformers 5.14.1 and safetensors 0.8.

Tests: new tests/test_inkling.py; Inkling in shared render/config parity and multimodal matrices (first audio modality kit); multimodal bridge uses renderer-specific turn-close tokens.

^{Reviewed by Cursor Bugbot for commit f235f2a. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add Inkling renderer with text, image, and audio support

• Adds renderers/inkling.py implementing InklingRenderer for the thinkingmachines/Inkling model, with chat-template parity, multimodal content parts (text, image, audio), tool declaration/invocation, reasoning effort formatting, and bridge/truncation support.
• Adds InklingRendererConfig in renderers/configs.py with a validated reasoning_effort field (named labels or float in [0.0, 0.99]) and an internal image_cache_max field.
• Adds parse_inkling in renderers/parsing.py to extract reasoning_content, text content, and structured tool calls from Inkling model outputs.
• Registers thinkingmachines/Inkling in the model-renderer map, modality support map (image + audio), and renderer registry in renderers/base.py.
• Risk: bumps the transformers minimum from >=4.50.0 to >=5.14.0 (first release with native Inkling processor support), which is a breaking install-time requirement change.

^{Macroscope summarized f235f2a.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR introduces a new multimodal renderer (InklingRenderer) with ~1100 lines of new code, new configuration, parsing logic, and a significant dependency version bump. New feature additions of this scope warrant human review.

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

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

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

^{Reviewed by Cursor Bugbot for commit 7883679. Configure here.}