fix(bdc): portfolio_investments() no longer counts holdings twice

[dgunning]

Bead: edgartools-6xxb

Problem

PortfolioInvestments.from_xbrl creates one row per InvestmentIdentifierAxis member. Those members have no hierarchy: there are no element labels and no domain-member arcs. So a filer's subtotals, and a second schedule's copies of its holdings, were read as holdings.

On every BDC measured, the rows summed higher than the filer's own undimensioned InvestmentOwnedAtFairValue (latest 10-K):

BDC	before	after
FSK	+48.4%	+11.47% ⚠︎
HTGC	+21.0%	+1.37%
CSWC	+18.9%	+0.46%
OBDC	+17.7%	+6.35% ⚠︎
ARCC	+15.7%	+0.28%
MAIN	+14.1%	−0.13%
TSLX	+5.9%	+1.74%
MSIF	+5.6%	0.00%
GBDC	+2.0%	+0.70%

MAIN's FY2025 10-K (0001396440-26-000016) summed to $6,296,964,000 against a reported $5,518,117,000. The MCP tool edgar_fund(bdc_portfolio) reported the inflated number.

Change

A row is excluded only when other rows account for its value exactly. Each rule comes from a real filing:

• Company totals such as "… and Total X" (HTGC). The sum is matched only against rows under the same heading.
• Parent rows that equal the sum of the rows extending their identifier:
  • MAIN: Bolder Panther Group, LLC | Secured Debt equals tranches 1.1, 1.2 and 1.3.
  • ARCC: Ivy Hill Asset Management, L.P. equals its instruments.
• One identifier spelled two ways with the same fair value (CSWC).
• Rows without a cost whose fair value equals a costed row. These are the affiliate roll-forward restating a holding under its own member name (OBDC Blue Owl Credit SLF LLC(c), FSK).

Currency: a fair value, cost or principal in the filer's reporting currency is no longer replaced by a foreign-currency fact. Before this, TSLX's Hippo XPA Bidco counted SEK 214.1M instead of USD 23.2M.

Reporting the remaining gap: OBDC and FSK still tag holdings in ways no exact rule recognises, so instead of guessing, the collection now exposes:

• reported_total_fair_value: the filer's own balance-sheet figure;
• reconciliation_gap: how far the rows' total is from that figure;
• excluded: the rows left out, each with a reason.

to_context() and the MCP bdc_portfolio result warn when the gap exceeds 2%. The MCP result also carries reported_total_fair_value.

I sampled the excluded rows by hand. They are company-level rows with no cost of their own (SPF Borrower LLC as the total of 4 loans, OPH NEP Investment, LLC as the total of 3), or explicit "Total" rows.

Verification

• New regression test: tests/issues/regression/test_6xxb_bdc_restated_rows.py, with 8 offline tests and 1 network ground-truth test.
  • It covers each rule, plus a silence check: a company named "Total Quality Logistics", a parent that is not the sum of its children, and two equal-valued costed tranches are all kept.
  • It checks that the most precise reported total wins (ARCC tags it at decimals −6 and −5), that the reporting currency beats a foreign one, and that a gap is surfaced rather than hidden.
  • The network test pins MAIN's reported $5,518,117,000, a gap under 0.2%, and 76 excluded rows.
• Updated test: tests/bdc/test_bdc_industry.py. Its synthetic "grouping subtotal" row is now excluded, and the test asserts that. The sector totals it pins are unchanged.
• Fast lane: 7,808 passed. The one failure, test_dt1f1_wrapped_item_headers (20-F lengths), also fails on main locally.
• BDC network tests, one file per call: test_bdc.py 236 passed, test_bdc_nonaccrual.py 20, plus the get_latest_bdc_report_year() treats a 200 as a complete report, and the 2026 one is missing ARES CAPITAL CORP #1146, AFFIRMATIVE_PATTERNS fails to extract MAIN non-accruals due to footnote wording change #835 and 6xxb files.
• Lint: ruff reports no new findings.

Not in scope:

• The DERA soi.tsv dataset path may have the same double count. It is a separate code path.
• The remaining OBDC and FSK gaps.

🤖 Generated with Claude Code