[dgunning]

Fixes #1505

Problem

_exclude_restated_rows (added in 5.60.0) drops a row as a parent when the rows extending its identifier sum to its fair value. It compares fair value only, so two real loans worth the same look like a header and its tranche. On ARCC's 10-Q 0001628280-26-050307, the $50.5M Proofpoint second-lien loan (cost $52.6M) was dropped as the total of ... loan 1 (cost $51.6M), along with two $100k pairs. The kept rows came to $29,298.6M against a reported $29,349.3M.

Fix

A row that states its own cost is kept when only one row extends it. The true one-row headers in this filing (APG, Bragg Live Food, Partnership Capital Growth III, ACAS Equity, European Capital UK SME Debt) state no cost and stay excluded.

Measured

16 BDC 10-Qs (ARCC, MAIN FY2025, and the latest from HTGC, FSK, OBDC, CSWC, GBDC, PSEC, BXSL, TSLX, GAIN, PFLT, NMFC, TCPC, CGBD and WHF): only ARCC changes, going from 1,444 to 1,447 rows, and its gap to the reported total goes from −$50.70M to $0.00M.

The reporter's second suggestion, requiring a costed multi-row parent to match its parts' summed cost, is left for 6.0. It moves MAIN's exclusions from 13 to 2.

Verification

• tests/issues/regression/test_issue_1505.py: an offline sibling pair is kept, an offline cost-less header is still excluded, and a network check that ARCC keeps Proofpoint and reconciles to the dollar. The first and third tests fail without the fix.
• test_6xxb_bdc_restated_rows.py, test_3vad_* and tests/bdc/test_bdc_mcp_portfolio.py pass.

Targets 5.62.0 as a cherry-pick onto 5.62.x: it is a 5.60.0 regression with no workaround.

🤖 Generated with Claude Code

[mmitc]

Thanks for the quick fix. It closes ARCC exactly: we measured 1,444 to 1,447 rows, gap -$50.70M to $0.00M, matching your numbers. We re-parsed 45 BDC filings live, on main at 2460109 with and without this change. That turned up one filer outside your 16-filing sample that the new rule moves the wrong way:

PhenixFIN (PFX), 10-Q 0001213900-26-085667, 2026-06-30: 67 to 83 rows, kept fair value $801.3M to $1,090.8M (+$289.5M), against PFX's stated $301.6M. PFX tags each company subtotal with the same fair value, cost and principal as its only holding, so "a costed row with one extending row" now keeps those subtotals as holdings. For example, Non-Controlled/Non-Affiliated Investments - Power Stop LLC equals its one term loan. Keeping them also stops Controlled Investments ($128.5M) and Affiliated Investments ($36.7M) from being recognised as totals.

NMFC FY2025 10-K (0001496099-26-000008): +$2.1M. ACI Parent's preferred is counted twice, as ACI Parent Inc. | Preferred shares and as its | Non-accrual status copy (same $2.087M fair value, same $20.124M cost).

The other moves are real loans coming back: ARCC Q1 -$4.0M to $0.0M, MSIF -$0.4M to $0.0M, and small additions at ARCC FY2025, BCSF, TRIN and BBDC.

A narrower rule that fixes Proofpoint without the PFX harm: when the parent and its parts all state cost, exclude the parent only if the costs match too. A cost-less parent stays excluded as today. Across the 45 filings this changes only Proofpoint. Adding "or a one-row parent's stated coupon differs from its part's" also restores Nordic Ferry and PumpTech, which share cost with their sibling.

Upstream tests can't tell these rules apart. Every variant gives identical results on the offline BDC tests (116), the regression suite (3,928), the fast suite and the live MAIN/FSK/OBDC/BCSF portfolio tests. A regression test on PFX 0001213900-26-085667 would pin it. Happy to send that as a PR if useful.

---

Generated by Claude Code