Add classification capability

[the-hercules]

What?

Adds a non-generative classification capability, for models that answer typed questions about a given state instead of generating content. Each answer is a probability, a chosen option, or a position on a scale, optionally with a confidence and a per-option or per-level probability distribution.

It follows the embeddings precedent: a capability enum value, a model interface, a builder, a result DTO, and lifecycle events.

Closes #296.

Why?

Today the only way to use a classification or judgement model is TEXT_GENERATION with a JSON response schema. That asks a generative model to print a number, and it throws away the calibrated confidence a classification model returns. It also means provider packages that support these models, such as Fueled/ai-provider-for-ollama#103, have to ship their own DTOs and entry points outside the client.

docs/REQUIREMENTS.md already lists classification as a non-generative feature the client must support.

How?

• CapabilityEnum::CLASSIFICATION.
• ClassificationModelInterface::classifyResult(array $state, array $questions): ClassificationResult.
• ClassificationQuestion, typed by ClassificationQuestionTypeEnum:
  • binary: a yes/no statement, with no criteria.
  • choice: criteria map each option key to a description, with at least two options. Keys must be non-integer strings, because PHP turns integer-like keys into integers, which can then encode as a JSON list.
  • score: criteria are an ordered list of level descriptions, lowest to highest, with at least two levels.
• ClassificationAnswer:
  • Records its question type.
  • Typed getters getProbability(), getChoice() and getScore() throw if called on the wrong type. getConfidence() and getProbabilities() are always available.
  • A score is a number on the scale, from 0 to levels − 1, and may fall between levels.
  • NaN and out-of-range values are rejected.
• ClassificationResult implements ResultInterface.
• ClassificationBuilder, created via AiClient::classify($state):
  • Picks a model through ModelResolutionTrait, as PromptBuilder does.
  • Validates state and question keys.
  • Checks that every question was answered and that each answer fits its question (type, option, scale, probability keys). Otherwise it throws a RuntimeException.
• BeforeClassifyEvent and AfterClassifyEvent go through the client's PSR-14 event dispatcher. The after-event fires only for valid results.
• README and docs/ARCHITECTURE.md updates.

Intentionally not included:

• No provider implementation; that belongs in provider packages.
• Details specific to Jev or Ollama stay in providers: the noul name, Ollama's limits (26 options, 64 questions, 64 KB body), the score legend, and optional true/false descriptions on binary questions. Each could be added to core later without a breaking change.
• No static one-call API yet, only the fluent builder.

Naming is open for maintainer feedback: CLASSIFICATION vs DECISION, AiClient::classify(), and the event names. See the shape proposal in #296.

Changelog Entry

Added - Classification capability for models that answer typed binary, choice, and score questions with probabilities and confidence, via AiClient::classify().

Use of AI Tools

AI assistance: Yes
Tool(s): Claude Code, [tool used for the initial draft]
Model(s): Claude Opus 5.5, [model used for the initial draft]
Used for: An initial implementation was drafted with AI. I then reviewed it with Claude Code against #296 and Fueled/ai-provider-for-ollama#103, which led to reworking the question and answer shape, adding answer validation and lifecycle events, and expanding the tests. All code was reviewed and tested by me.

🤖 Generated with Claude Code

[github-actions]

The following accounts have interacted with this PR and/or linked issues. I will continue to update these lists as activity occurs. You can also manually ask me to refresh this list by adding the props-bot label.

If you're merging code through a pull request on GitHub, copy and paste the following into the bottom of the merge commit message.

Co-authored-by: the-hercules <[email protected]>
Co-authored-by: juanlentino <[email protected]>

To understand the WordPress project's expectations around crediting contributors, please review the Contributor Attribution page in the Core Handbook.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 87.78%. Comparing base (77a5995) to head (7ba63a8).

Additional details and impacted files

@@             Coverage Diff              @@
##              trunk     #300      +/-   ##
============================================
+ Coverage     86.58%   87.78%   +1.20%     
- Complexity     1383     1503     +120     
============================================
  Files            69       75       +6     
  Lines          4449     4872     +423     
============================================
+ Hits           3852     4277     +425     
+ Misses          597      595       -2     

Flag	Coverage Δ
unit	87.78% <100.00%> (+1.20%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[jeffpaul]

Is "classification" the right/best naming option here (versus "decisioning" or something else)?