Issue #31: radius-jewel framework

[jonatanferm]

Summary

Generic mechanic for "a jewel socketed in a tree socket modifies passive nodes inside its radius" — the foundation that Timeless jewels (#30) and Cluster jewels (#21) both build on. This PR ships the generic infrastructure plus end-to-end identification + application of vanilla node-modifying radius jewels (e.g. 10% increased Maximum Life to nearby allocated passives on a Crimson Jewel).

• pob_data::jewel_radius: canonical 3.16+ radii table (Small=960, Medium=1440, Large=1800, Very Large=2400, Massive=2880) plus the 3.15-and-older fallback, with a tree-version picker that mirrors PoB's setJewelRadiiGlobally.
• pob_engine::jewel_radius: node-position math (group + orbit + orbit_index → Cartesian), nodes_in_radius / allocated_nodes_in_radius for scanning, identify_radius_jewel for detecting vanilla radius jewels via mod text, and apply_radius_jewels for the per-allocated-node mod replay. The radius-suffix stripper ( to nearby allocated passives, from Passives in Radius, …) lets mod_parser mint canonical names (Life, Strength) instead of long-form aliases.
• Character::socketed_jewels (SocketedJewels keyed by tree-socket node id) round-trips through CharacterSnapshot. perform::init_env_with_bases calls apply_radius_jewels after item application; each in-radius allocated node receives one mod copy sourced as Source::Passive(target_node) so the breakdown attributes the bonus to the in-radius node.
• Cluster / Abyss / Timeless / Charm jewels share the same storage but are filtered out by identify_radius_jewel's subtype check — they keep their own dispatch paths in Cluster jewel sub-graph support #21 / Timeless jewels — keystone/notable replacement within radius #30.

Deferred to follow-ups (tracked by their own issues):

• Timeless jewel keystone / notable substitution (Timeless jewels — keystone/notable replacement within radius #30).
• Cluster jewel sub-graph synthesis (Cluster jewel sub-graph support #21).
• Named-unique handlers (Watcher's Eye, Healthy Mind, Karui Heart, Pure Talent, Intuitive Leap, Conqueror's Efficiency). The HandlerKind enum on RadiusJewel enumerates the dispatch surface they'll plug into.
• PoB-XML round-trip for <Slot name="Jewel <NodeId>" itemId="...">.

Test plan

• cargo fmt --all -- --check clean.
• cargo clippy --workspace --all-targets -- -D warnings clean.
• cargo test --workspace — 302 tests pass (was 291 baseline).
• 11 unit tests in pob_data::jewel_radius + pob_engine::jewel_radius cover the radii table, position math, scan, identification (Cluster/Abyss/Timeless skipped, non-radius rare jewel skipped, explicit ring sizes), suffix stripping, and end-to-end mod emission.
• 7 integration tests in crates/pob-engine/tests/radius_jewels.rs exercise the framework against the real 3_25.json tree fixture, including a regression that empty-alloc compute is byte-identical to baseline (no drift when no nodes are in radius).
• End-to-end test verifies socketing 10% increased Maximum Life to nearby allocated passives into a real tree socket plants an Inc Life mod sourced from the in-radius allocated passive in the modDB.

🤖 Generated with Claude Code

[jonatanferm]

[loop-status: lgtm] Foundation for the radius-jewel mechanic — Timeless (#30) and the named-unique handlers (Watcher's Eye etc.) build on this. Right architecture: pob_data::jewel_radius ships the canonical radii table with tree-version picker, pob_engine::jewel_radius exposes the math + scan + identification + replay; Character.socketed_jewels round-trips through CharacterSnapshot. Each in-radius allocated node receives a Source::Passive(target_node) attribution so breakdown shows the right contributing chain. HandlerKind enum is the right dispatch surface for unique-named handlers landing later. Conflict with #190 was a clean dual-storage situation (cluster's Character.jewels alongside radius's Character.socketed_jewels) — kept both since they target different mechanics; the PR body itself flags that the two will eventually unify. CI green post-rebase, +18 tests including end-to-end against the live 3.25 tree with a regression for empty-alloc baseline drift.