feat: EDGAR 10-K → fiber fragment adapter (SP-ADAPT-TREE-001, real edition)

[mdheller]

What

Moves fibered retrieval from fixtures onto real documents. edgar_fiber_adapter.py fetches a company's latest 10-K from EDGAR (data.sec.gov submissions → primary doc), extracts its Item-level table of contents (the tree descend navigates), and emits a Crystal Atlas fragment fiber_projection.project consumes: document root ⊑ Item sections, registrant anchored under Item 1 to the real filing URL. No model, no datastore.

Proof — it runs on a REAL filing

The committed fixture edgar_aapl_10k.fragment.json is Apple's actual latest 10-K (17 real items, anchored to the real sec.gov URL), and the test descends its real table of contents to "Item 1A — Risk Factors" with the keyword navigator — no network in the suite.

Honest scope

The Item extractor is a heuristic over filing HTML (strip tags → match Item N. Title, dedup) — it recovers the standard Part I–IV items, not a production parser (exhibits/XBRL R-files out of scope). It emits only what the filing structurally states — no ownership %s or subsidiary edges are invented. Cross-fiber ownership (Exhibit 21 → GLEIF → E_R edge, which enables /multi-hop) is the documented follow-up. Network lives only in the fetch path; parsing + fragment emission are pure and offline-tested.

Tests (+3, full suite 301 green)

extract_items on a compact HTML fixture · to_fragment anchors the registrant under Item 1 · the real Apple fragment projects and descend navigates it. Stdlib-only (urllib); the live fetch path is exercised by the CLI (--cik 320193).

Unblocks the --model eval: point the real Claude scorer at these real fragments (needs the key, which the CI env lacks).