docs(langsmith): document trajectory variable and trajectory online evaluators - #5873

Closes DOC-1598.

Why

Trajectory online evaluators score the flat, deduplicated message list for a session, including the tool calls and tool results that the thread message variables leave out. The docs described only all_messages, human_ai_pairs, and first_human_last_ai, so there was nothing covering the trajectory variable or how a trajectory evaluator differs from a thread evaluator.

What changed

src/langsmith/prompt-template-format.mdx

• Adds trajectory as a fourth evaluator variable, with its resolved message shape: role of human/ai/system/tool, and content as either a string or a list of content blocks (text, reasoning, tool_call).
• Notes that trajectory is mutually exclusive with the thread variables, and that LangSmith rejects a prompt combining them.
• Adds an "Example with trajectory context" showing a template and the input LangSmith provides.

src/langsmith/online-evaluations-multi-turn.mdx

• Adds an "Evaluate agent trajectories" section: what the judge receives compared to a thread evaluator, creating an evaluator with {{trajectory}} in the prompt, where feedback is written, the evaluator kind being fixed at creation, and the supported and rejected settings.

Please review carefully

• Timing. GA is 2026-09-11, so this should not merge before that date. The pages read as generally available, with no beta tag.
• Two claims from the design doc were deliberately dropped. The doc says the UI "prevents or flags" a prompt that mixes trajectory with thread variables, but I could only find backend validation (schemas.py:_validate_trajectory_variable_mapping); the frontend adds trajectory to the available variable paths and nothing more. The pages therefore say LangSmith rejects the combination without attributing it to the UI. The trajectory_evaluators_enabled flag name is also omitted.
• Limits. I did not claim the existing thread-level limits (500 threads per execution, 7-day run age) apply to trajectory evaluators, since trajectory resolution runs through its own paging path. Worth confirming with engineering whether they do, and whether trajectory evaluators count against the 10-per-workspace cap.
• Message shape. This resolves the ticket's open question. Verified in smith-go/runs/v2/messages/conversation_types.go, the {"trajectory": item["messages"]} evaluator input in smith-backend/app/models/run_rules/rule_application/trajectory_eval.py, and the CRUD constraints in smith-backend/app/schemas.py. The example input matches the sample data shown in the evaluator config UI.
• Trace scope. The design doc says both thread and trace scopes are supported. The creation flow I found is the Threads source, so the docs describe threads only. Flagging in case trace-scope trajectory evaluators need coverage too.

Checks

• make lint_prose clean on both files.
• make broken-links reports no broken links.

AI disclosure

Drafted with Claude Code (Opus 5), reviewed by me. Product behavior was verified against the langchainplus implementation rather than taken from the design doc.