feat: AgentControl data model, parsing & interpolation (AIC-2662)

[ctawiah]

Requirements

• I have added test coverage for new or changed functionality
• I have followed the repository's pull request submission guidelines
• I have validated my changes against all supported platform versions

Related issues

• Implements AIC-2662 — Step 2 of the Java AI SDK (epic AIC-2629).
• Foundation (AIC-2661, chore: scaffold LaunchDarkly Java AI SDK module foundation (AIC-2661) #170) has merged; this PR is now based on main.

Describe the solution you've provided

Step 2 of the Java AI SDK: the AICONF data model and the defensive JSON-protocol parsing layer. No LDAIClient methods, tracker, or evaluation (those are AIC-2663+).

Scope note: Mustache interpolation has been pulled out of this PR. Per the SDK team's supply-chain guidance we will not link the external com.samskivert:jmustache artifact. The Mustache engine will be vendored (source copied into an internal, relocated package) together with the Interpolator in AIC-2695, which gates the v1.0 release (AIC-2666). This PR therefore has no third-party templating dependency.

Public data model — com.launchdarkly.sdk.server.ai.datamodel (immutable, builder-based, documented):

• LDMessage (Role enum + content), ModelConfig (name/parameters/custom), ProviderConfig, ToolConfig (root-level tool), JudgeConfiguration (+ nested Judge), and the AIConfigMode enum.

Internal parsing — com.launchdarkly.sdk.server.ai.internal (excluded from published Javadoc/sources):

• LDValueConverter — depth-capped (MAX_DEPTH = 100) LDValue → plain Java tree. Integral numbers within ±2^53 decode to Long, otherwise Double.
• AIConfigParser + AIConfigFlagValue — LDValue → strongly-typed parse. Malformed/missing/wrong-typed fields never throw; tools fall back to model.parameters.tools[]; evaluationMetricKey resolves to the first non-blank of evaluationMetricKey / evaluationMetricKeys[]; _ldMeta scalars are boxed to preserve tri-state enabled (absent ≠ false).

Design decisions (documented in code)

• Data-model placement: kept in ...server.ai.datamodel rather than extracting a shared module now, to avoid premature abstraction; can relocate if a client-side AI SDK ever needs it.
• Boxed vs primitive: _ldMeta scalars (enabled/version) are boxed in the parsed value so "absent" is distinguishable from a concrete value; resolved configs (later step) will use primitives with documented defaults.
• Number precision: whole numbers outside ±2^53 cannot be represented exactly and decode to the nearest Double.

Additional context

Field shapes and behaviors mirror the Python (python-server-sdk-ai) and JS (js-core/packages/sdk/server-ai) reference SDKs for parity. 22 unit tests cover defensive parsing fallbacks and number/tool/metric-key resolution; full module build (compile + checkstyle main/test + javadoc + tests) is green.

---

Note

Low Risk
New library surface and internal parsers only; no client evaluation or runtime wiring yet, with broad unit test coverage for malformed payloads.

Overview
Adds the Java server AI SDK data model and defensive flag-variation parsing for AI Configs (AIC-2662). Mustache interpolation is not in this PR; build.gradle documents vendoring Mustache in AIC-2695 instead of linking jmustache.

Public API — New LDAIConfigTypes in datamodel groups immutable nested types: Mode, Message (+ Role), Model, Provider, Tool, and JudgeConfiguration (+ Judge), with builders where needed and Message.withContent reserved for future interpolation.

Internal parsing — AIConfigParser maps LDValue → AIConfigFlagValue without throwing on bad input; LDValueConverter turns nested JSON into plain Java maps/lists with depth cap (MAX_DEPTH = 100) and integral-number handling within ±2^53. Parsing covers _ldMeta (boxed enabled for absent vs false), tools (root tools over model.parameters.tools[]), and evaluation metric key (scalar then first non-blank in evaluationMetricKeys[]).

Build — Javadoc no longer sets failOnError = false now that public types exist; empty sonatype* entries removed from gradle.properties. Unit tests cover parser and converter edge cases.

^{Reviewed by Cursor Bugbot for commit 69416f0. Bugbot is set up for automated code reviews on this repo. Configure here.}

[jsonbailey]

Approved from me but would be good to also get Todds review.

[mattrmc1]

nit: For the .NET SDK, we opted to organize shared shapes like Message, Model, Provider, Tools, etc into a static class LdAiConfigTypes. This reduces the number of files and frees up generic class names (ie Message) within the namespace. Could do the same thing here for consistency

[tanderson-ld]

Is it the case that none of the enums added in this PR are for public consumers to switch over? The reason I bring that up is adding another value to an enum is breaking in java.

[ctawiah]

There is a high probably that we can introduce a new enum for the mode in the future. The mode is mainly for internal use and theres no plans to expose that to customers.