fix(serve): rank resolve_element ties by closeness, not the alphabet

[jfrench9]

Summary

resolve_element scores matches in tiers (exact, prefix, token, label, qname), and a phrase like "revenue" prefixes hundreds of concepts in one tier. Ties fell to fact count and then the alphabet. A filing's fact counts usually separate them; a taxonomy loaded in 0.17.0 reports nothing, so on US GAAP "revenue" answered RevenueChangeInJudgment and "Revenue Commissioners, Ireland [Member]" first and ranked us-gaap:Revenues past 100. This ranks ties by closeness instead, and scores a label that differs from the phrase only by punctuation or a plural as exact.

Changes

serve/tools.py (surface)

• Within a score tier: concepts with more facts first (unchanged), then non-abstract before abstract, then numeric before non-numeric (an amount before a text block, a policy or an enumeration), then the shortest concept name, then the qname.
• A phrase whose squashed form (case, spacing, punctuation removed, one trailing s ignored) equals the concept's name or preferred label scores 90, the exact tier. "operating lease liability" now matches the label "Operating Lease, Liability"; "cash and cash equivalents" matches US GAAP's singular standard label "Cash and Cash Equivalent".

Output Impact

INTERNAL for serialization: holon, TAVI and model output are unchanged. The order of resolve_element results changes wherever a tier ties on fact count. That is always the case on a taxonomy and occasionally on a filing. Scores, match counts and match fields are unchanged. Anything that measures resolve_element rankings, such as the financial analysis benchmark, can see different first matches when run against this version.

Testing

just test-all via the commit hook: ruff, format, basedpyright clean; 545 passed, 2 skipped. The new test (test_resolve_element_ranks_a_taxonomy_with_no_facts_to_lean_on) fails on main (RevenueChangeInJudgment first) and passes here.

Rank of the expected concept, before → after, over 19 queries per source:

Query	US GAAP 2025 (taxonomy)	NVDA 10-K 0001045810-26-000021
revenue → Revenues	129 → 1	1 → 1
operating lease liability → OperatingLeaseLiability	18 → 1	12 → 1
cash and cash equivalents → CashAndCashEquivalentsAtCarryingValue	16 → 1	4 → 1
share-based compensation → ShareBasedCompensation	206 → 1	34 → 1
inventory → InventoryNet	44 → 1	2 → 1
earnings per share → EarningsPerShareBasic	2 → 1	1 → 1
stockholders equity → StockholdersEquity	16 → 1	1 → 1
research and development → ResearchAndDevelopmentExpense	18 → 1	1 → 1
net income → NetIncomeLoss	5 → 2 (ProfitLoss first)	1 → 1
accounts receivable → AccountsReceivableNetCurrent	26 → 5 (AccountsReceivableNet first)	1 → 1

The other queries (revenues, cash, assets, goodwill, long-term debt, income tax expense, depreciation, operating income) were 1 before and stay 1 on both sources, except that NVDA does not report us-gaap:Cash. A qname query (us-gaap:Assets) ranks 1 on every source. After the change, every query on AAPL 10-Q 0000320193-26-000020 ranks the expected concept 1 wherever the filing reports it, and all six GASB 2026 exposure-taxonomy queries rank 1 (pension liability was 14 before).