Add investment parsing support for 14 BDC tickers

[HaCk3Dq]

Adds structured investment identifier parsing for TSLX, MSDL, GSBD, MFIC, BCSF, PFLT, PSBD, RWAY, LIEN, PNNT, BCIC, GECC, OFS, and TRIN.

Includes regression coverage for company names, investment types, XBRL field anchors, aliases, facilities, series, and rollups. Parser tests: 157 passed.

[dgunning]

Thanks for this — and for coming back to the same subsystem after #747. The BDC label formats are a genuinely nasty surface and the breadth here is welcome.

I checked it by running the parser over two BDCs the PR doesn't mention, on main and on this branch, and diffing the (identifier, company_name, investment_type) triples. Good news first, then one blocker.

The classification gain is real

	investments	Unclassified → typed	typed → Unclassified
MAIN	751	120	0
OBDC	640	11	0

Nothing regresses in the direction of losing a type. That's a solid improvement on issuers you weren't targeting.

Blocker: company_name regresses on 13 investments

Same run, company_name changes on 13 investments across those two BDCs — and each change is wrong. Your 157 tests pass because none of them cover MAIN or OBDC labels.

Pattern 1 — the industry is taken as the company (12 of 13, all OBDC). On labels shaped <Company> | <Industry> <type> investment | <relationship>:

label: Wingspire Capital Holdings LLC | Specialty finance equity investment | Affiliated
  main: company='Wingspire Capital Holdings LLC'   type='Unclassified'
  PR:   company='Specialty finance'                type='equity'

The type does get classified, which is the win — but the company name is replaced by the industry. Also hits AAM Series 1.1 Rail and Domestic Intermodal Feeder, LLC, AAM Series 2.1 Aviation Feeder, LLC, Amergin Asset Management, LLC, Blue Owl Cross-Strategy Opportunities LLC, Fifth Season Investments LLC, LSI Financing 1 DAC, LSI Financing LLC and Wingspire Capital Holdings LLC (several appear twice under ... 1 and ... | Affiliated variants).

A related one on the same BDC, where a leading relationship fragment gets eaten:

label: Controlled/affiliated - debt commitments, First lien senior secured revolving loan
  main: company='Controlled/affiliated - debt commitments'
  PR:   company='debt commitments'

Pattern 2 — a share-class prefix is taken as the company (1, MAIN).

label: DTE Enterprises, LLC | Class AA Preferred Member Units (non-voting)
  main: company='DTE Enterprises, LLC'
  PR:   company='Class AA'

That one looks connected to the new 'Class AA units' entry in INVESTMENT_TYPES matching inside Class AA Preferred Member Units.

This is the failure mode we treat as most serious: no exception, no empty field, just a confidently wrong company name flowing into PortfolioInvestment.company_name.

Reproducing

from edgar.bdc import find_bdc
inv = find_bdc("OBDC")[0].portfolio_investments()
print({i.company_name for i in inv if "Specialty finance" in i.identifier})
# main: the eight real company names
# this branch: {'Specialty finance'}

One decision to confirm, not necessarily blocking

525 of MAIN's 751 investment types change, mostly dropping a tranche suffix — 'Secured Debt 1' → 'Secured Debt', 'First lien senior secured revolving loan 1' → '... revolving loan'. I think that's the right normalization, since identifier still disambiguates the tranches, but it silently changes results for anyone grouping by investment_type. Please confirm it's deliberate; if so it wants a CHANGELOG note under [Unreleased].

What would land this

1. Preserve the company span on the two label shapes above.
2. Add MAIN and OBDC labels to tests/test_bdc.py. The current 157 cases are organised per new ticker, so a cross-issuer regression has nowhere to fail — the 13 above went unnoticed for exactly that reason. A handful of parametrized cases from already-supported issuers would close the gap.
3. Confirm the tranche-suffix normalization.

Everything else looks good: no ticker symbols hardcoded in the parser, no existing test modified or deleted, and all 228 tests in tests/test_bdc.py pass on the branch. It's the cross-issuer blast radius that needs containing, not the approach.

🤖 Review assisted by Claude Code

[HaCk3Dq]

Thanks, fixed in the latest push.

Preserved the company span for the OBDC pipe-delimited and comma-delimited label shapes.
Added parametrized MAIN and OBDC regression cases, including taxonomy member context.
Confirmed that removing trailing numeric tranche suffixes from investment_type is intentional. The complete label remains in identifier, and this behavior is now asserted by a test and documented under [Unreleased].

All 161 investment identifier parser tests pass.

[dgunning]

Approving. The blocker is fixed, and I verified it the same way I found it — parsed MAIN and OBDC on this branch and on the merge-base (2df53a2c, not current main, so the diff isolates the PR) and compared the (identifier, company_name, investment_type) triples.

The 13 wrong company names are gone

	investments	Unclassified → typed	typed → Unclassified	company_name changed
MAIN	751	120	0	0
OBDC	640	13	0	0

The classification gain is intact and OBDC's went 11 → 13. Your reproduction case now returns the real names:

['AAM Series 1.1 Rail and Domestic Intermodal Feeder, LLC',
 'AAM Series 2.1 Aviation Feeder, LLC', 'ABF - Specialty finance',
 'Amergin Asset Management, LLC', 'Blue Owl Cross-Strategy Opportunities LLC',
 'Fifth Season Investments LLC', 'LSI Financing 1 DAC', 'LSI Financing LLC',
 'Wingspire Capital Holdings LLC']

The DTE Enterprises, LLC / Class AA case is fixed by the right mechanism too — adding 'Class AA Preferred Member Units' to INVESTMENT_TYPES so the longer type wins the match, rather than special-casing the label.

Tranche normalization is confirmed deliberate and documented in [Unreleased], and the measured split matches the description: MAIN 405 suffix drops + 120 classification gains, OBDC 62 + 13, no unexplained churn.

One thing I fixed rather than sent back

7dbe51f6, pushed to this branch. The new descriptor-pipe branch matches case-insensitively and returned the captured span verbatim, so the filer's typography became the value: OBDC writes "Specialty finance equity investment", which produced 'equity' (11) and 'debt' (2) — the only lowercase-initial types across either BDC, sitting beside 'Preferred Equity', 'Common Equity' and 'Secured Debt' from every other branch. Grouping by investment_type then split the concept in two, which is the property your CHANGELOG entry sells. Title-cased at the single site that can produce it, with a parametrized test over three label casings.

Re-measured after that change: 13 values recased, no company_name touched, no type gained or lost, no lowercase-initial types left. tests/test_bdc.py is 236 passing.

One residual, not blocking

The four new parametrized cases cover exactly the four labels I reported. That closes those, but the original problem was that 157 tests passed while nothing covered MAIN or OBDC at all — four labels is not coverage of those issuers, so the next cross-issuer regression still has nowhere to fail. Worth a follow-up that pins a slice of real MAIN/OBDC triples, rather than growing this PR.

Thanks for the careful turnaround on this one — the diagnosis and the fixes were both right.

🤖 Review assisted by Claude Code