refactor(characters): extract shared preparation-context pipeline

[QMalcolm]

Closes #126

What

The preparation-context derivation (class lookup → mode → DAG evaluation → cap → max level → pool config → eligible pool) existed twice: tolerantly in build_prep_state_map (read path) and strictly in activate's with-chain (write path), with opt_activation_cap / pool_eligible existing solely to soften the strict versions.

Now one shared pipeline in two stages:

1. prep_class_and_mode/3 — active class + preparation mode; cheap, no DAG evaluation.
2. prep_context/4 — resolved cap, max prepared level, pool config, eligible pool. Fallible steps stay tagged ({:ok, _} | {:error, _}) so activate pattern-matches them strictly while preparation_state degrades tolerantly.

opt_activation_cap and pool_eligible are deleted; both ad-hoc context map shapes are gone (one carried a pool_name key its consumer never read).

Why two stages

activate must reject a wrong preparation mode before any DAG evaluation. My first single-stage version evaluated eagerly — and the existing mode-rejection test caught it, because its fixture deliberately has no graph, encoding exactly that ordering contract. The split preserves original error precedence (no-class → wrong-mode → cap → pool → eligibility → cap-limit) and avoids wasted evaluation on mode errors.

Verification

No behavior change: full suite passes unchanged (311 + 55 tests), credo, dialyzer, CodeScene delta all green.

Summary by Bito

• Refactored the character preparation pipeline into a shared, two-stage process to unify logic between read (preparation_state) and write (activate) paths.
• Introduced `prep_class_and_mode/3` to perform a lightweight check of the preparation class and mode before expensive evaluation.
• Introduced `prep_context/4` to centralize the resolution of activation caps, pool configurations, and eligible items, ensuring consistent behavior across the system.
• Improved error handling in the activation path by leveraging the structured context returned by the new pipeline.