SOTA convergence manifest 2026-09-23 with its citation review and the generator fixes it found - #100

Summary

This adds the dated convergence manifest catalogs/sota-convergence/manifest-20260923.json, which the 2026-09-23 cross-family lane run will use. It changes no ledger row; rows are re-recorded only through that later lane run, and manifest-20260922.json is untouched.

How it was produced

1. extract_layers.py at main 168a3a8: 20 foundation layers, 12 trading layers, 154 trading entries, 342 star candidates.

2. github_freshness.py at full coverage: 469 repositories, 0 errors.

3. The saved sota-convergence workflow (wf_4069d065-fd4, 184 agents):

   • three Opus lanes: foundation, trading, and beyond-stars/awesome-lists;
   • two verification lenses per proposal (evidence and gap-fit), with two refuters;
   • a completeness critic.

   It made 90 proposals and dropped 9. No agent was lost.

4. An independent citation review (wf_75979946-4da, three Opus evidence reviewers, Read/Grep only). It found 36 findings: 3 high, 16 medium, 17 low. Seven are generator defects.

5. The generator defects are fixed in build_manifest.py with 12 regression tests:

   • the lane-to-card join now uses the repository slug, so card URLs ending in /tree/… or /releases/tag/… no longer lose their lane review;
   • citations into the checkout keep repository-relative paths;
   • pins that were never compared are published as not_compared / null instead of false;
   • OS-package pins get distro_managed status;
   • a bogus upstream tag is flagged instead of published.

6. The manifest was rebuilt with the review as a per-row overlay. No generated text was hand-edited.

Result

• 116 confirmed components; 95 before the join fix.
• 25 unique components are behind upstream. Their pin qualification receipts come from SOTA refresh evidence wave: pin qualification, observability comparison, hardware profiles #87 and Gap resolution 2026-09-22: 40 receipts and a 315-gap ledger across all 32 layers #81.
• 20 pins could not be compared.
• 46 candidates: 15 not_adopted_confirmed, 15 refuted_keep_but_compare, 7 refuted_targeted, and the rest unverified.
• The critic found 6 layers with no confirmed selection and 12 gaps with no candidate. The gaps include broker-state failures, IBKR acceptance, pre-2020 survivorship and point-in-time news.
• Citation overlay: 29 findings attached, 83 rows flagged, 0 left general.

Known limits, from the Opus verification (accept)

• Checkout-relative citations resolve only when the manifest is built from the checkout the lanes cited.
• In two trading rows (execution-broker and one other), the lane review is still lost because of a catalog_id mismatch, not the URL join.
• The overlay's component-field match can attach one finding to more rows than the reviewer meant; this is disclosed.
• Rebuilding 20260922 with these fixes would change only the pin and upstream fields: 20 rows false → null, and 6 rows get latest_flag. The committed 20260922 file is left as it is.

Checks (local, merged with main 796f759)

• python3 -m unittest: 2,507 tests OK, 338 skipped.
• validate.py, evidence_manifest.py --check, landscape.py, build_verdicts.py --check, validate_convergence.py --all-recorded, build_ecosystem.py --check, trading_gates.py --check and component_matrix.py --check: all pass.
• Guarded gitleaks on the range: no leaks.

🤖 Generated with Claude Code