Refresh the SOTA convergence manifest from the 2026-09-22 layer-verdict lane run - #66

What

Refreshes catalogs/sota-convergence/manifest-20260922.json from the saved Claude workflow sota-convergence (run wf_2388074a-e82: foundation, trading and beyond review lanes, two Opus refuters per proposal, one completeness critic; 110 Opus 5/high children, 908 requests, 1,064,906 output tokens, 46.5 min) over GitHub metadata for 469 repositories fetched 2026-09-22 14:57 UTC.

The Codex lane could not run: the ChatGPT account's usage limit is exhausted until 2026-09-28 (thread 01a0c9a3-102b…, no purchase). Cross-family review of this PR is therefore not established; the ledger rows stay pending_lanes for the PR-5 owner to record with codex_absent. The PR-5 owner cherry-picks this commit into PR-5; this PR stays open unmerged until PR-5 merges and then closes with the reference.

Tooling (tools/sota-convergence/build_manifest.py), each with regression tests

• Review-lane grouping ids outside the taxonomy go to lane_groupings instead of fake trading rows (which broke scripts/landscape.py coverage).
• Every lane's entry for a (layer, repository) is kept: the row takes one by a documented precedence (adversarially verified beats unverified, the owning lane beats beyond, then lexical), the rest are published under other_lane_reviews; an entry naming another card is never folded in.
• A refuted demotion/unmaintained proposal resolves to the component's own baseline class (confirmed_default / confirmed_conditional / confirmed_selected), never an unconditional confirmed_default.
• Adversarial verdicts are shared across layers for the same lane, repository and proposal kind, with a provenance evidence line.
• Lane entries key by catalog_id when set, so two cards for one repository in one layer (alpaca-py execution and data) keep distinct rationales.
• Sanitization keeps the file basename and JSON pointer of redacted citations (<work-dir>/file#pointer, <host-path>/file#pointer), never republishes a home-directory name, scrubs session-scoped /tmp paths and bare session UUIDs; assert_no_leak refuses the scripts/validate.py session-identifier pattern.
• --citation-review overlay stamps an independent review's findings onto the rows they name (citation_review) and keeps the rest under citation_review.general, with counts; generated text is never hand-edited.
• 66 new tests (124 in tests/test_sota_convergence.py).

Independent review

evidence/artifacts/sota-convergence-review-20260922/manifest-citation-review.json (hash-listed) records the four-reviewer Opus pass over the first refresh: one reviewer on the tooling diff, three opening every cited item behind why_selected / comparison_that_would_overturn on 157 component rows. Of its 36 findings, 5 were generator defects (fixed above, then re-reviewed in a build → review → fix → recheck round, wf_299a1543-f6e, with the two recheck residuals resolved by the coordinator and covered by tests) and 31 are citation-quality findings on lane text (line-number drift, execution claims without receipt paths, unfalsifiable or prescriptive overturn statements, uncited star-audit claims) now attached to their rows. Pins and upstream blocks of all reviewed rows matched manifests/stack.json, the trading cards and github-freshness.json.

Evidence classes

Metadata and source review by the lanes; adversarial verification inside the workflow; local integration checks below. No install, run, benchmark or provider trial. The beyond lane's 9 proposals beyond the per-lane cap of 20 are recorded as unverified, not as refuted.

Checks (worktree, after rebase on d093a4c)

python3 -m unittest tests.test_sota_convergence   # Ran 124 tests OK
python3 -m unittest                               # Ran 1438 tests OK (skipped=204, system python)
python3 scripts/validate.py                       # 69 components / 1746 hashed files / 139 receipts
python3 scripts/landscape.py --root .             # passed
python3 scripts/validate_catalogs.py              # passed
python3 scripts/validate_convergence.py --all-recorded --root . --json   # valid
python3 scripts/validate_foundation.py --root . --json                  # ok, 20 layers / 54 decisions
python3 scripts/catalog_decisions.py --check      # passed
python3 tools/sota-convergence/build_verdicts.py --check --root .       # checked
python3 scripts/build_ecosystem.py --check        # passed
git diff --check                                  # clean
gitleaks git . --config .gitleaks.toml --max-target-megabytes 2 --log-opts=d093a4c..HEAD   # first head: no leaks (explorer HTML skipped by size: incomplete coverage); amended head: local guarded scan blocked by another session's scan lock, hosted secret-scan is the check of record

Private inputs, sealed lanes.json (sha256 315946d7…) and per-child usage are under the private state directory named in docs/tasks/2026-09-22-sota-convergence.md (agent-lab).

🤖 Generated with Claude Code