Parse AI behavior actionParams into typed models

[mocha06]

Motivation

BehaviorInput.action_params was a raw dict | None, and every consumer re-walked it with dual-alias branches (aiBehaviorParams/ai_behavior_params, actionsAttributes/actions_attributes): the model validator, the SDK preflight and pipe/phase-transition checks, and MCP helpers such as the pipe-ID extraction. That duplicated dict-walking contradicts the repo's "parse, don't validate" / self-guaranteeing-types rules and multiplies as more fields are added to the payload.

Outcome

action_params is now a typed Pydantic model (AiBehaviorActionParams → AiBehaviorParams → action and capability attributes) with per-declared-field camelCase aliases matching the GraphQL input schema and extra="allow" so unknown/sibling keys pass through verbatim. Every consumer parses once and reads typed attributes, so no dual-alias branches remain in SDK or MCP agent code paths.

Behavior is unchanged for every payload that was already valid end-to-end. snake_case inner keys — which the API always rejected while the toolkit's dual-alias branches accepted them — are now normalized to the declared wire names, so that input works for the first time. Aliasing is strictly per declared field (never a blanket snake→camel pass, which would corrupt genuinely snake_case siblings like card_id/to_phase_id), and action metadata stays a pass-through dict so GET→update round-trips serialize identically.

Tested e2e a few times

No new validation rules land here — capability and provider hardening build on this typed base.

Closes #416.

[adriannoes]

Summary

Solid prep PR for typed behavior actionParams.

Nested models with per-field aliases replace the dual-alias dict walks without changing valid end-to-end wire payloads, and capability/provider hardening correctly stays out of scope for #424. Review ran checkout plus local tests (2922 passed, ruff green); CI lint and test are green.

What worked well

Per-field camelCase aliases with extra="allow" keep AutomationActionParams siblings (card_id, to_phase_id) intact while normalizing declared AI-behavior keys. Lenient BehaviorPayload reads versus strict BehaviorInput writes is a clean split, and the slug-resolve path deep-copies before parse so caller metadata is not mutated through pydantic references.

Also noted

• The public resolve_and_populate_field_refs helper early-returns without rewriting keys when no slug fetch runs, while populate_referenced_field_ids now only reads declared camelCase names. In-repo MCP/CLI/facade paths already model_dump(by_alias=True) first, so they stay correct. If that helper remains a supported entry point for raw snake_case dicts, normalizing on every exit (or documenting dump-first) would match #416's "normalize at the entry edge" wording.