Extend is_stopword check in weighted_edit_similarity to either side of a cluster

[pudo]

Motivating case

From the harness corpus (contrib/name_comparison/cases.csv):

name1	name2	label	quality	category	current score
Bashar xal-Assad	Bashar al-Assad	true	STRONG	Fat Finger Typo	~0.67

Below the 0.7 alert threshold despite being a STRONG match.

Diagnosis

After tokenisation on the hyphen, the residue stage clusters xal ↔ al (0.51 overlap rule fires on the shared al). Lengths 3 and 2 give a tiny per-side cost budget; one edit blows past it and the cluster scores ~0, dragging the weighted average from above-threshold down to 0.67.

The cluster-weight policy in nomenklatura/matching/logic_v2/names/distance.py already has a stopword down-weight that would solve this:

if len(match.qps) == 1 and len(match.rps) == 1:
    if is_stopword(match.qps[0].form):
        match.weight = 0.7

al is in resources/text/stopwords.yml (line 21). But the lookup only checks qps[0] — i.e. xal, the typo'd query side — so the result-side particle never marks the cluster as low-information.

Proposed fix

Two small changes:

1. nomenklatura/matching/logic_v2/names/distance.py — make the stopword check fire when either side is a known stopword:

   if is_stopword(match.qps[0].form) or is_stopword(match.rps[0].form):
       match.weight = 0.7

   Rationale: a cluster is low-information if the intended token (either side) is a connective particle. The matcher shouldn't require that the noisy side also happens to land on a known particle.

2. resources/text/stopwords.yml — audit particle coverage. Already present: al, de, la, el, del, du, le, da, der, von. Worth adding for Arabic/Persian/Hebrew/Romance/Germanic/Celtic naming particles commonly seen in screening data: bin, ibn, ben, abu, abd, dos, mac, mc, van. (Curated, not exhaustive — precision over quantity per repo conventions.)

The existing 0.7 weight is already calibrated; this just ensures it triggers in the symmetrical case.

Out of scope

• New is_name_stopword / NAME_PARTICLES resource — the existing is_stopword + resources/text/stopwords.yml already model the right concept.
• Particle insert/delete across the full alignment (e.g. Mohammed bin Salman vs Mohammed Salman) — that's symbol pairing's job, not residue distance.
• Tokenizer changes (e.g. collapsing hyphens) — too broad, would affect unrelated cases.

Validation

The harness in contrib/name_comparison/ exercises this end-to-end:

• The Bashar case (Bashar xal-Assad vs Bashar al-Assad) should rise above 0.7.
• Run the full suite to confirm no regressions on cases.csv labelled outcomes (especially STRONG-tier negatives that share particles).