fix: a unique price column is not a foreign-key target

[kiyeonjeon21]

Why

Relationship inference treated any unique, non-null column as a valid foreign-key target. A small price list has unique prices, so every money column whose values happened to coincide got "explained" as a foreign key into it. Measured at 81% / 68% / 54% on a realistic fixture when the hard eval dataset was first built - which is how this was found.

The damage is worse than a wrong edge in the graph. Both endpoints of a relationship are excluded from the semantic model, so a phantom edge silently deletes the table's real measure:

before:  Sale:      measures = sales_count             <- sum_line_amt is gone
         PriceList: measures = price_list_count        <- sum_list_amt is gone
after:   Sale:      measures = sales_count, sum_line_amt
         PriceList: measures = price_list_count, sum_list_amt

Nothing in the output says why. On the hard dataset this removed sum_net_amt from the invoice table entirely.

What

A non-id target is credible only when the foreign column names it outright (nameSimilarity >= STRONG_NAME_SIMILARITY), which is what a genuine natural key looks like: sku -> sku, region_cd -> region_cd. Uniqueness alone no longer qualifies a column as a join target.

isIdLike moves to signals.ts as the single shared definition. The two halves of the engine had quietly disagreed: semantic-model.ts already carried the comment "a unique, non-null column like amount is a real measure, not a key" - and inference did not know that.

Verification

• Isolated fixture (a 4-row price list whose prices coincide with a sales table's line amounts): the phantom edge disappears, the legitimate sku -> sku join survives at 100%, and both measures come back. A regression test pins exactly this.
• npm run eval:engine 18/18 and npm run eval:engine:hard 25/25, both unchanged - so no real edge was lost, including the three natural-key joins the hard dataset depends on.
• npm run check exit 0; npm run test:cli 142 pass / 0 fail

ROADMAP step 6.1. Found by building a harder eval dataset, which is the second engine defect that work has surfaced.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CKZ9UVgw2Mhiu6T5YAG7k6