Multi-label entities: use max similarity across all names

[claude]

Summary

• Replaces Node.name: str with Node.names: list[str] so entities can carry multiple names (e.g. "Meridian Technologies" and "Meridian Tech")
• Name similarity seeding in propagate_similarity now computes max(soft_tfidf(a, b)) across all name pairs, ensuring the closest match is always used
• IDF computation includes all names from all entities
• Graph I/O serializes the names list and loads legacy single-name format for backward compatibility

Context

Prerequisite for progressive merging (#25): when entities merge during propagation, the merged entity retains all names from both sides so subsequent name-similarity seeding remains accurate.

Test plan

• New test: multi-label entity seeds similarity from best name pair
• New test: all names contribute to IDF computation
• New test: multi-label names survive save/load round-trip
• New test: legacy single-name JSON format loads correctly
• All 58 existing tests pass unchanged

Closes #27

🤖 Generated with Claude Code

[claude]

One issue worth fixing: backward-compatibility shim for legacy single-name JSON format.

load_graph (graph.py:52-56) adds fallback logic to handle the old "name" key:

raw_names = node_data.get("names") or [node_data["name"]]
names=raw_names if isinstance(raw_names, list) else [raw_names],

There are no old-format graph files in the repo and no external users. Per CLAUDE.md conventions ("never add backward-compatibility shims... Refactor completely"), this should just read node_data["names"] directly. The test_load_legacy_single_name_format test should be removed as well — it tests the shim.

Everything else looks correct: the max-over-name-pairs seeding, IDF over all names, names[0] for functionality phrase pairs (same semantics as the old .name), and the union-find display. Clean refactor otherwise.

[monneyboi]

Remove the backward-compatibility shim in load_graph — just read node_data["names"] directly. Delete test_load_legacy_single_name_format as well.

[claude]

Looks good. All 57 tests pass. The Node.name → Node.names refactor is applied consistently across graph.py, match.py, and all tests. Name similarity seeding correctly takes the max over all name pairs, IDF includes all names, and the backward-compat shim was properly removed per CLAUDE.md conventions. No dead code or half-finished refactors.