feat(trajectories): add exact token content attribution

[bradhilton]

Agent: Unattributed

Add TokenFlag.CONTENT and sequence-level content_attribution_complete. Mark caller/model body tokens only after rendering exactly reproduces every engine-consumed prompt ID and the sampled output IDs. Consumers must require complete attribution and a nonempty mask before averaging.

The recorded support matrix covers streaming/nonstreaming Chat Completions, Messages and Responses for ART-registry Qwen3, Qwen3.5/Qwen3.6 dense/MoE models, OpenPipe/Qwen3-14B-Instruct with its documented tool-history limit, and GPT-OSS 20B/120B. Gemma 4, DeepSeek V4, explicit Base variants, raw Completions, custom templates, unknown models, multi-generation Responses and incomplete/divergent metadata leave CONTENT unset.

Historical validation: 100 trajectory/renderer-parity tests and 386 backend tests passed, with three skips; repository hooks and adversarial review passed. A disposable H200 vLLM 0.24.0/Qwen3-0.6B check verified exact prompt IDs across all three APIs and both response modes. CI covered the locally unavailable Megatron/TE partition. Live SGLang parity and a dedicated cached optional-renderer CI job remain absent. Caladan #82 supplies the companion engine prompt-token carriers.

Draft. Responsible agent and current readiness are unverified; historical evidence is not a fresh qualification.

Historical agent ownership is unverified; this remains in the unattributed bucket.

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
openpipe-art	🟢 Ready	View Preview	Jul 22, 2026, 1:22 AM

💡 Tip: Enable Workflows to automatically generate PRs for you.

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
openpipe-art	🟡 Building	–	Jul 22, 2026, 1:22 AM

💡 Tip: Enable Workflows to automatically generate PRs for you.