fix(text): aligned tables, cover-only form detection, 8-Ks read with their exhibits

[jfrench9]

Summary

Comparing the two servers on Workiva's latest earnings 8-K showed three problems in xbrlkit's text tools. search_text on the 8-K found nothing in the release, because an 8-K loaded as its cover page alone. The financial tables put values in the wrong column. And the release, loaded by its own URL, came back labelled a 10-K. This PR fixes all three. The table fix also reaches the RoboSystems sec server, which renders filing text through xbrlkit.

Changes

text/tables.py (table → markdown)

• When a $, ( or ($ merged into its number, the cell it sat in was dropped, so every value after it in the row moved one column left. The same happened when a trailing ) or % was attached. Cells now keep their positions. Each column a symbol stood in is joined to the next column when no row has values in both. That also lines up Workiva's layout, where a value without a $ spans the symbol and number columns.
• A self-closing <td ... /> (Apple's filing agent writes spacers this way) was read as an open tag and swallowed the cell after it. It now reads as an empty cell.

serve/session.py (cover form)

• _form_on_the_cover searched the first 400,000 characters for "Form 10-K", so a press release's safe-harbor citation made it a 10-K. Only the top 2,000 characters are searched now. Real covers name their form within about 200.

serve/ (8-K exhibits)

• An 8-K or 8-K/A loaded from EDGAR, on both the XBRL path and the document-only path, now fetches its EX-99 exhibits through other_documents / read_other_document. They are joined into one text: ## Form 8-K, then ## EX-99.1 and so on, each a section with its offset. This is the layout the RoboSystems sec server reads an 8-K in. If the exhibits can't be reached, the 8-K loads on its own as before.
• join_texts and FORM_SECTION_ID are added to the declared xbrlkit.serve surface, so a host that builds the text itself can use the same assembly.
• The items_note, next steps, server instructions, the documents tool description and serve/README.md now point to read_text / search_text when the exhibits are in the text and the text tools read it whole. Under pure, or when only the tagged text is read, they still point to documents, because there the exhibits aren't in the text the tools search.
• If one exhibit can't be read, the others still go into the text. Two exhibits both typed plain EX-99 each get their own section id.

Output Impact

CHANGED OUTPUT in the text tools (search_text, read_text, describe_filing sections), not in the serializations. Holon, TAVI and model.json don't go through the table renderer.

• Tables: every filing's markdown tables change shape. Spacer columns are gone and values sit under their headers. Character offsets into the text change with them, so an offset saved from an earlier version no longer points at the same place.
• 8-Ks from EDGAR: the text now includes the EX-99 exhibits, and sections.items gains form_8k / ex_99_1. A load costs the index fetch plus one fetch per exhibit.
• Documents loaded by URL or path: a document whose only "Form X" mention is below its first 2,000 characters now reports form: null instead of a guess.
• RoboSystems: it renders through build_text, so it picks up the table change when its pin moves. Text it has already cached or indexed keeps the old rendering until rebuilt.

Testing

• The pre-commit hook ran ruff, basedpyright and the suite on each commit; the last run had 696 passed, 2 skipped. New tests cover the $ position shift, the colspan-2 layout, a symbol column that holds real values, self-closing cells, cover vs. citation, join_texts, an 8-K read with its release, an 8-K whose exhibits can't be reached, one exhibit failing while the others load, the exhibit guidance in pure and tagged-text modes, and a 10-K that doesn't reach for exhibits.

• Tables, old vs. new renderer: five documents, counting rows with values outside the columns most rows use. The total fell from 1,574 to 329:

  • Workiva Q2 2026 EX-99.1 (0001445305-26-000059): 47 → 0
  • Apple 10-K (0000320193-25-000079): 79 → 5
  • Microsoft 10-K (0001193125-26-323660): 270 → 27
  • NVIDIA 10-K (0001045810-26-000021): 111 → 11
  • JPMorgan 10-K (0001628280-26-008131): 1,067 → 286

  I read the six tables this count flagged as worse; each is now correct. For example, Apple's buyback total sits under "Approximate Dollar Value", and JPMorgan's loan maturities sit under their buckets.

• Cover form: the first "Form X" match is at 125–215 characters in the four 10-Ks and the Workiva 8-K, and at 13,455 in the Workiva release.

• 8-Ks live from EDGAR:

  • WK 8-K (0001445305-26-000059, inline XBRL cover): EX-99.1 is read in, and "free cash flow|guidance" returns 20 hits, all labelled EX-99.1.
  • A 2018 Workiva 8-K with no XBRL (0001445305-18-000131, document-only path): EX-99.1 is read in, with 7 hits.

🤖 Generated with Claude Code