docs(evaluate): missing section for rubric_based_multi_turn_trajectory_quality_v1 + Rubric.type / criterion.rubrics requirements

[zettaittenani]

Problem

docs/evaluate/criteria.md is missing documentation for the rubric-based multi-turn evaluator, and the implicit requirements around Rubric.type and criterion-level rubrics aren't explained for any of the rubric-based criteria. Each gap causes a real, easy-to-hit failure in practice.

Gap 1: No section for rubric_based_multi_turn_trajectory_quality_v1

The metric exists upstream — registered in metric_evaluator_registry.py (PrebuiltMetrics.RUBRIC_BASED_MULTI_TURN_TRAJECTORY_QUALITY_V1) and implemented in rubric_based_multi_turn_trajectory_evaluator.py — but the corresponding section in criteria.md is missing. The other two rubric-based metrics (rubric_based_final_response_quality_v1, rubric_based_tool_use_quality_v1) have full sections with EvalConfig examples; this one doesn't.

User impact: someone trying to evaluate multi-turn trajectory quality with custom rubrics has no guidance, and falls back to copying the single-turn section, which uses subtly different semantics.

Gap 2: Rubric.type field semantics aren't documented

Every RubricBasedEvaluator subclass declares a RUBRIC_TYPE ClassVar that acts as a filter — only rubrics whose Rubric.type matches that value are picked up:

Metric	Required Rubric.type (from source)
rubric_based_final_response_quality_v1	FINAL_RESPONSE_QUALITY
rubric_based_tool_use_quality_v1	TOOL_USE_QUALITY
rubric_based_multi_turn_trajectory_quality_v1	TRAJECTORY_QUALITY

None of these values appear anywhere in criteria.md (or anywhere else in adk-docs based on a code search). The existing JSON examples in criteria.md don't set the type field either, so users following the examples and then attaching per-case rubrics via EvalCase.rubrics end up with type=None rubrics that the evaluator silently drops.

User impact: rubric scores come back empty / None with no clear error — only a Rubric ... not found in the rubrics provided to the metric. warning if the user happens to be reading log output closely.

Gap 3: criterion.rubrics is required, EvalCase.rubrics is additive — neither is stated

RubricBasedEvaluator.__init__ asserts that criterion.rubrics is non-empty (rubric_based_evaluator.py:332). create_effective_rubrics_list then unions criterion-level rubrics with per-case rubrics (filtered by type). The docs don't mention either fact:

• The current criteria.md examples put all rubrics on the criterion, so the assert never fires for users who copy them verbatim.
• But users who try to attach rubrics on individual EvalCase entries (a natural pattern when each case has case-specific rubrics) and leave criterion.rubrics = [] hit AssertionError: Rubrics are required. at init with no docs-level explanation of why.
• Similarly, users who put rubrics on both expect criterion rubrics to be overridden per case; they're actually unioned.

User impact: confusing assert at init, or unexpected "extra" rubrics being scored per case.

Proposed direction

A single docs PR that:

1. Adds a ## rubric_based_multi_turn_trajectory_quality_v1 section to criteria.md, modeled on the existing rubric_based_final_response_quality_v1 section (with a TRAJECTORY_QUALITY example), and the corresponding row to the criteria table at the top of the page.
2. Adds a small "Notes On Rubrics" subsection under each of the three rubric-based metric sections that states the required Rubric.type and the criterion-vs-EvalCase rubric relationship.
3. Adds the explicit type field to the existing JSON examples in the two existing rubric-based metric sections so users have a complete template.

No code changes — docs/evaluate/criteria.md only.

Filing this as an issue first per the CONTRIBUTING guideline for "New Documentation". Happy to follow up with a draft PR for review if maintainers think the direction is right.