Fix: dupe-date-price heuristic no longer drops distinct parcels of a multi-parcel deed

[mividtim]

Problem

run_heuristics's flag_dupe_date_price keyed duplicate detection on (sale_date, sale_price) alone (optionally prefixed by jurisdiction). Distinct parcels conveyed together in a single multi-parcel deed share a sale date and price — and per-parcel records frequently carry the deed total — so every lot but one of a subdivision sale was flagged as a "duplicate" and (when drop=True) dropped. This silently discards most comparable sales of a subdivision, which is especially damaging for vacant-land / rural models that depend on those comps.

Fix

A genuine duplicate is the same parcel reported more than once at the same date and price (duplicate data entry / shell trade); distinct parcels that merely share a date and price are not. The fix appends the parcel key to the duplicate signature when a key column is present, so only true same-parcel repeats are flagged. Null keys fall back to the unique per-sale key_sale so they never collide.

Behavior change (deliberate, not a no-op)

Multi-parcel-deed lots that were previously dropped now survive, increasing the sale count feeding downstream ratio studies and models. The previous behavior was over-aggressive; the new behavior is correct.

Scope

Fixes flag_dupe_date_price (heuristic #2) only. flag_dupe_deed_date (heuristic #1) intentionally flags multi-parcel deeds by deed_id + sale_date — that is its purpose (multi-parcel-sale detection), distinct from duplicate-report detection — so it is left unchanged.

Tests

Adds tests/test_sales_scrutiny_heuristics.py: distinct parcels in one deed survive; a same-parcel repeat is still dropped; one parcel sold twice on different dates is kept; the jurisdiction-scoped branch composes correctly. Backward-compatible: only applies when a key column is present (always the case for hydrated OpenAVMKit sales).

Note for reviewer

The duplicate signature is built by string concatenation to match the sibling flag_dupe_deed_date heuristic's existing style; happy to switch to a groupby(...).transform("size") form if preferred.

[github-actions]

Thank you for your contribution.
Please sign our CLA at the following link:
Click here to sign the CLA.

No action is required from you in this PR thread. Once you have signed the CLA externally, a maintainer will verify your signature and record it here on your behalf by commenting:

---

I affirm that this contributor has signed the CLA

---

_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}

[nikhilsi]

A data point in support of this PR and #351 together. We applied both on
top of 7952236 and ran them against another county's data: roughly 130K
parcels with about 24K sales in the modeling window. The (date, price)
dupe heuristic had been dropping around 12% of those sales as false
duplicates; with the parcel key from this PR they are retained, and the
orphaned bulk deeds that then surface are caught by the #351 scrutiny
check rather than leaking through. The ratio-study COD improved on the
larger sample. The two PRs complement each other and we would suggest
reviewing them as a pair.