Record the landscape sweep's measured usage (stopped run no longer 'unknown'; completed run's effort verified)

catalogs/sota-convergence/manifest-20260923.json

@@ -17460,10 +17460,10 @@
    "Discovery: one Sonnet 5 researcher per layer at effort max (WebSearch <=10, WebFetch <=4, gh api <=30 per layer); adversarial verification: a facts/identity refuter (Sonnet 5, max) and a fit/standing refuter (Opus 5.5, max) per layer, each defaulting to refuted; a candidate survives only when neither refutes it; one Opus 5.5 (max) completeness critic and one bounded follow-up round over at most 8 critic-flagged layers.",
    "Source review and GitHub metadata only: no candidate was installed, run, benchmarked or compared with a winner; survival means the proposal withstood fact and fit checks, not that the candidate is better than a winner.",
    "upstream_now values are as observed by the researcher's gh api calls during the run (2026-09-23/24 UTC), not a github_freshness.py fetch.",
-   "Call counts cover the discovery agents only; refuter and critic calls are not counted here. Completed run wf_38aa6d5d-d6c: 119 agents (32 discovery, 31+31 refuters, 1 critic, 8 follow-up discovery with 8+8 refuters), 12,123,864 subagent tokens, 2,823 tool uses, 6,615 s wall clock, as reported by the workflow runtime.",
+   "Call counts cover the discovery agents only; refuter and critic calls are not counted here. Completed run wf_38aa6d5d-d6c: 119 agents (32 discovery, 31+31 refuters, 1 critic, 8 follow-up discovery with 8+8 refuters), 12,123,864 subagent tokens, 2,823 tool uses, 6,615 s wall clock, as reported by the workflow runtime; per-request usage from child-usage.mjs (a different counter, not to be summed with it): all 119 children resolved to claude-sonnet-5 or claude-opus-5-5 at effort max, 4,225,386 output / 3,544 input / 138,862,781 cache-read / 9,733,523 cache-creation tokens (retained at evidence/artifacts/landscape-sweep-20260923-attempts/child-usage-wf_38aa6d5d-d6c.json).",
    "The completeness critic's follow-up round (8 critic-flagged layers, same two-refuter rule) is merged into this lane with the first round.",
    "Prior records of survivors (no adopt or reject decision, so they remain candidates): PaddlePaddle/PaddleOCR is discovery_only in catalogs/landscape/claude-independent-discovery.json, where a 2026-09-21 refuter recorded PP-StructureV3 below MinerU2.5 on OmniDocBench; anthropics/claude-code-action is targeted_candidate (3 votes) in catalogs/sota-convergence/sdk-runtime-coverage-20260922.json with an executed CI smoke arm (/gated_items/rows/0), and decision HOST-09 (docs/harness-rules-convergence-20260922.md) keeps it unadopted until a pinned-commit check of the action repository is run.",
-   "A first run (workflow wf_28d47bbc-1b5) at lower worker effort (discovery Sonnet high, refuters Sonnet medium and Opus high) was stopped after 9 of 32 discovery returns and no refuter returns, when the effort policy changed to max for every worker. It is retained, not merged, at evidence/artifacts/landscape-sweep-20260923-attempts/wf_28d47bbc-1b5.json with each returned repository, its call counts, and whether this run re-proposed it; its provider usage was not captured (unknown).",
+   "A first run (workflow wf_28d47bbc-1b5) at lower worker effort (discovery Sonnet high, refuters Sonnet medium and Opus high) was stopped after 9 of 32 discovery returns and no refuter returns, when the effort policy changed to max for every worker. It is retained, not merged, at evidence/artifacts/landscape-sweep-20260923-attempts/wf_28d47bbc-1b5.json with each returned repository, its call counts, and whether this run re-proposed it. Its usage, measured afterwards from the retained transcripts by agent-lab's .claude/workflows/child-usage.mjs: 17 discovery children (9 returned, 8 stopped in flight), claude-sonnet-5 at effort high, at least 168,052 output / 302 input / 8,233,147 cache-read / 773,809 cache-creation tokens. The tool reports the run incomplete, so these are lower bounds and the unobserved remainder is unknown; its output is retained at evidence/artifacts/landscape-sweep-20260923-attempts/child-usage-wf_28d47bbc-1b5.json.",
    "Post-review amendments (Codex cross-family review of #153, 2026-09-24), votes and survival unchanged: RD-Agent's comparison now scores on a final segment its feedback loop never sees; both TruffleHog comparisons use a controlled, revocable credential for the verification arm; MinerU's stale v1.0-era OmniDocBench figures and CJK framing are removed from its gap and comparison. Final review of #153: PaddleOCR's rank-1 OmniDocBench claim is corrected to 3rd (README at f133a71e9e) and its proposed label lowered from targeted_candidate to keep_but_compare, because its fit vote's text called the rank-1 basis void while recording refuted=false; claude-code-action's comparison cites its prior records and its Release evidence is corrected; RD-Agent keeps both stay conditions."
   ]
  },