Grand catalog finalization (2026-09-23): live board #140

Live board for the user's 2026-09-23 goal:

• Every layer's lanes are resolved with evidence at the current landscape (the 20260923 SOTA manifest).
• The GitHub workflow is finalized.
• A release exists that a new PC installs from. Install and finalization run on the new PC.
• Evidence from future hosts and runtimes flows back in.
• Then the next north-star stages.

Every session updates this issue with comments or edits. The dated snapshot is agent-lab #29, owned by curate-harness-rules-convergence (agent-lab-e9); it links here once the release is cut.

Current state (2026-09-25 ~19:40Z, workstation coordinator session)

Owners are Claude sessions on the workstation (nativestack-5975wx-20260925) unless another host is named. The user's current focus: clean up the noise, then work the grand catalog and the foundation layers behind the north-star actions (token efficiency, memory/RAG, local models end to end).

Account limits: the Opus weekly limit resets 2026-09-30 21:00 ET, so subagents run on Sonnet until then. Codex resets 2026-09-30 23:50 local.

User-only items: rotate the Alpaca paper keys (HTTP 401 blocks paper runs); sign in to IB Gateway for paper; name an off-host backup destination; provide a Tavily key; approve any GPU window that stops nativestack-generation.

Exit criteria for "finalized"

• All 32 layers are recorded in a non-grandfathered wave (20260923) with sealed cross-family lanes, and 0 are pending_lanes.
• verdict-review-gate is a required check on main (Add the required verdict-review-gate: sealed cross-family evidence for every changed verdict row #135 35416786; ruleset 23739774 confirmed by GET, 2026-09-23 19:25Z).
• Linux platform_status derives from registered, independently reviewed host receipts (platform_status()), with no declared upgrades.
• Release v2026.09.23.1 is cut, attested, verified and re-pinned in adoption/manifest.json. The "not in the pinned release" notes are removed. (2026-09-25: the current release is v2026.09.25.1, re-pinned by Point new machines at the attested release v2026.09.25.1 #259; the re-record release after the wave is still due.)
• The new-host grid and component matrix are regenerated on the release commit.
• The readiness snapshot (agent-lab ci: bump actions/setup-python from 6.3.0 to 7.0.0 #29) is finished and links this issue.

Critical path (in order)

1. Add the required verdict-review-gate: sealed cross-family evidence for every changed verdict row #135 verdict-review-gate: DONE. It was squash-merged as 35416786, pinned to the head bac1600b that agent-lab-e9 spot-checked.
   • Ruleset 23739774 was PUT from origin/main:.github/main-ruleset.json. The GET after it confirms:
     • enforcement active;
     • required checks validate, token-report, secret-scan, dependency-review, osv-scanner and verdict-review-gate;
     • rules code_scanning, deletion, non_fast_forward, pull_request and required_linear_history;
     • an exact match with the committed file.
   • Open branches cut before 35416786 must merge main to report the new check: Record scheduled memory learning and daily native lifecycle evidence #116 (commented). Live record of paper runs: NautilusTrader JSON log, decision stream, read-only manifest #138 has been rebased (2d43fc69) and now reports the check. The IBKR branch will be rebased before its PR opens.
   • After a retarget, close and reopen the PR or push a commit. The edited run checks out the stale merge commit, as measured on [throwaway] Live retarget test of the #135 H1 fix; do not merge #143 (runs 35907554093, 35907626904 and 35908075608).
   • Make verdict-review-gate apply the landscape validator's withheld-key policy #144 (64863f48): the gate now imports the base's scripts/landscape.withheld_packet_keys instead of a copy that had drifted. Keys added by blind-lanes round 2 (gap_receipts, gap_receipts_note) apply with no gate edit. It passed independent review.
   • The re-record PR is the gate's first real use. It gets a read-only dry-run before it opens.
2. The 20260923 re-record of all 32 layers, owner agent-lab-17.
   • Inputs first:
     • Integrate gap wave 2 (agent-lab-17): 23 layers, 189 receipts, owner ledger #132 (gap wave 2), Blind both layer-verdict lanes for the 2026-09-23 re-record #141 (blind lanes) and Let lane packets carry each component's registered receipts #142 have merged.
     • The blind-lanes round-2 PR (claude/blind-lanes-round2-20260923) is under independent review. It changes the trust inputs lane_packets, build_verdicts (type-strict check_frozen_rows, which closes a Add the required verdict-review-gate: sealed cross-family evidence for every changed verdict row #135 residual), codex_lane and lane-provenance, so it must land before the re-record.
   • Run:
     • candidate set manifest-20260923 (SOTA convergence manifest 2026-09-23 with its citation review and the generator fixes it found #100); freshness was 469 repos, 0 errors;
     • lane_packets --withhold-labels, then blind_checkout --export;
     • the Claude and Codex lanes, then counterbalanced two-family adjudication;
     • record_verdicts --run-id 20260923.
   • The 2 pending_lanes layers (execution-broker, observability-hosting) are re-judged with the SOTA refresh evidence wave: pin qualification, observability comparison, hardware profiles #87 and Establish adaptive-paper-broker-trial on a main-built 1x paper trial (paper rung ready) #128 evidence. If they still split, they stay pending_lanes with the executed comparison named.
   • The PR registers the wave, rewrites all 32 rows and regenerates the matrix and grid. The 20260922 wave stays frozen. ETA 6-8 h after its inputs land. Add the required verdict-review-gate: sealed cross-family evidence for every changed verdict row #135 is dry-run against the branch before the PR opens.
   • After the re-record merges: a separate one-line PR widens ENFORCED_PLATFORMS (scripts/landscape.py:51) to {"macos-arm64", "linux-wsl2-x86_64"}.
     • It has to come after the re-record. Widening it today fails main on grandfathered 20260922 rows, for example foundation/ci-supply-chain zizmor Linux accepted against evidence that supports at most conditional.
     • landscape.py is a gate trust input, so it lands alone. The re-record changes no trust file.
3. The WSL2 accepted drop (about 50% to about 30%) is expected, and there is no owner. After Verdict-pipeline integrity before the 20260923 re-record: per-wave CI, unbiased packets, two-family rule, survivorship, receipt-backed status #122/Verdict integrity round 2: bind rows to sealed waves, strict single-lane authorization, immutable waves, sealed refutations, retained packets #124, accepted reflects evidence strength. All 26 Linux conditional winners are local_integration (15) or synthetic (11). accepted needs either an independently reviewed native_proven pass from a second physical machine, or a native_proven/measured_comparison winner that cites a registered non-lane evidence/ file. Mass "it runs" receipts from the authoring host will not be recorded, because that would be the bias the user wants avoided. The status rises two ways: (a) the re-record classes existing native receipts correctly, for example native-tavily-cli-20260920.json, native-dagu-retry-tests-20260920.json and native-headroom-mcp-20260920.json; (b) new hosts record reviewed use-stage receipts with --second-physical-machine.
4. Release v2026.09.23.1: superseded by v2026.09.24.1. One new-PC step failed as written (corrected 2026-09-24 ~06:15Z).** It is superseded by v2026.09.24.1 (step 4b).
   • Failing step: step 4's Claude context-mode install, claude plugin marketplace add mksglu/context-mode@6f0cc684… --scope user (recipes/README.md:125, docs/token-session-handbook.md:194, docs/token-efficiency-stack.json:360 at the tag), exits 1. A Claude marketplace source takes a branch or tag, not a commit. The evidence is 2026-09-24 community sweep: effort-guard SessionStart notice, working plugin-install commands, dated dispositions #161's isolated run, evidence/artifacts/community-sweep-20260924/plugin-marketplace-refs.json.
     • Workaround on .1: claude plugin marketplace add mksglu/context-mode --scope user, then check gitCommitSha as in main's step 4a.
     • Main 4a4c8a2 (2026-09-24 community sweep: effort-guard SessionStart notice, working plugin-install commands, dated dispositions #161) fixes the documents.
   • Also: the no-git archive check in bootstrap.md step 0 isn't bound to the release. It accepts an attested archive of another commit saved under the release name, which was reproduced. The default clone + tag + commit path is unaffected. Bind the no-git archive check to the release; report new-machine content drift in release_due; run step 6 under Python 3.13 #163 fixes it.
   • Found by agent-lab-87's readiness sweep and verified at source by github-automation-gap-closure.
   • Readiness verdict (agent-lab-87's Opus/max critic over 8 areas at main 4a4c8a28, 06:25Z): ready with workarounds, not fully ready. A new Linux/WSL2 PC can start from .1 with these workarounds:
     • step 0 by clone + tag, not the no-git archive;
     • context-mode by the ref-less add, then the gitCommitSha check;
     • step 6 as uv run --no-project --python 3.13 python scripts/adoption_status.py …, because plain python3 3.12 exits 2 (reproduced; fixed in Bind the no-git archive check to the release; report new-machine content drift in release_due; run step 6 under Python 3.13 #163 e76eb1a);
     • --claude-bin for install_claude_profile.py when claude isn't on PATH;
     • fix agent-lab's .mcp.json host paths by hand (fixed since, in agent-lab Reconcile foundation catalogs with current upstream and Claude evidence #49 7917b40d).
   • Verified sound on main:
     • the ruleset matches its file with no bypass actors;
     • 20/20 checks are green, and all 110 uses: are SHA-pinned;
     • tag attestations verify;
     • all 38 referenced paths exist at the tag.
   • User-only: macOS acceptance on a real Mac, and the native sign-ins.
   • User decision (2026-09-25 00:15Z): Codex usage is exhausted on the new ChatGPT account. Jobs fail with "try again at Sep 30th, 2026 11:50 PM" (agent-lab-e9). Codex PR reviews and every Codex lane are unavailable. That includes wave 20260924's two-family adjudication of any layer where the lanes disagree; the Codex lane itself finished beforehand, 32 of 32 clean. The options are the user's: add credits, sign in with another account, or wait until 2026-09-30.
   • Cut by agent-lab-e9: tag d1357e2b, which points to 40828dfe (all 20 checks green). publish-catalog run 35930806614 succeeded, and the release is immutable.
   • Re-pin Point new machines at the attested release v2026.09.23.1 #148 was squash-merged as 54f89a93 after an independent review found no issues. The review, from a fresh download:
     • both attestations verify (the archive under SLSA; the SPDX file under --predicate-type https://spdx.dev/Document);
     • verify-asset passes for both assets;
     • the archive holds 5472 files, the same as the tag tree;
     • release_due --strict reports current.
   • Step 0 checked from a fresh clone of main (2026-09-23 ~23:06Z):
     • release_due.py reports current with due: [];
     • the checkout lands on v2026.09.23.1, at the pinned commit 40828dfe;
     • the 16 formerly missing install paths are now in the release, including install_claude_profile.py, apply_claude_settings.py, adoption/mcp/claude-user.json, the launchd agents and embed_acceptance.py.
   • Start here: adoption/bootstrap.md from step 0, then adoption/platforms/<macos-arm64|linux-wsl2>.md. The native sign-ins (claude, codex, gh) happen on the new machine at step 3.
     4b. Release v2026.09.24.1: DONE. New-PC setup starts here (updated 08:07Z). The tag is annotated and points at 81385fe6. Re-pin Point new machines at the attested release v2026.09.24.1 #171 merged as 8e386e38, and main pins it.
   • It carries:
     • 2026-09-24 community sweep: effort-guard SessionStart notice, working plugin-install commands, dated dispositions #161 (the working plugin-install commands and step 4a's gitCommitSha check);
     • Cap each guarded job's CPU where the user manager delegates cpu #158 (the per-job CPU quota);
     • Bind the no-git archive check to the release; report new-machine content drift in release_due; run step 6 under Python 3.13 #163 (the release-bound archive check, and release_due reporting new-machine content drift, so a re-pin must pin main's new-machine files).
   • After the cut: publish, attestation verify with --source-digest/--source-ref, verify-asset, the re-pin PR, then an independent fresh-clone verification by github-automation-gap-closure.
   • Merge hold: CLOSED after re-pin Point new machines at the attested release v2026.09.24.1 #171 merged; tracked-file PRs can merge again.
   • Independent fresh-clone verification (github-automation-gap-closure, verify_release.sh): PASS.
     • Main pins .24.1, and release_due --strict reports current (0 due, 0 changed).
     • Step 0 checks out the tag at the pinned commit.
     • The release is immutable. verify-asset passes for both assets, the archive attestation is bound to refs/tags/v2026.09.24.1 and 81385fe6, the SBOM attestation (SPDX predicate) is bound to the commit, and a wrong digest is refused.
     • The archive holds 5631 files, equal to the tag tree.
     • The no-git pin one-liner reads the new pin, and the release-pin and docs tests on main pass (50 OK).
   • (history) The rule while the hold was active: don't merge any PR that touches a file release_due tracks (62 files plus README's Start here section on main). A change landing in that window fails the re-pin's --strict check.
     • Open PRs that touch tracked files (at 81385fe6; manifests/evidence.json alone is exempt): Blind lanes round 2: sanitized export, receipt aliases, two-family adjudication #145, Carry the dated manifest's surviving newcomers into every packet (--manifest-newcomers) #151, Adoption hygiene: native Claude Code pin is a floor (2.1.281), ai-memory assistant capture opt-in, ignore Claude Code local state #159, Record NativeStack workstation acceptance and retained failures #160, Require a reviewed use-stage pass for an accepted platform status #164, Saturation tracking loop: hash-chained per-layer ledger, weekly model-free report, one tracking issue #166, Mover v3 research plan and protocol draft from the verified 2026-09-24 SOTA sweep (not frozen) #167, Close the Alpaca native-fault engine gaps (C05 cancel-again sends DELETE; C04 documented definitive refusal) #169, Sync runtime-target IBKR entry and the grand-dashboard checkpoint with merged evidence (records only) #170. The list of tracked files is release_due.new_machine_files() on main. Local peer sessions were told at the go.
     • Adoption hygiene: native Claude Code pin is a floor (2.1.281), ai-memory assistant capture opt-in, ignore Claude Code local state #159 either merges before the tag or waits for the re-pin.
     • Record NativeStack workstation acceptance and retained failures #160 stays unmerged until its author fixes it, which puts it outside the window. agent-lab-e9's 72-agent review is on the PR:
       • 46 of 66 receipts are supportable, and 20 need changes;
       • the check source for content-assertion receipts isn't bound at catalog_revision 8eacd8b;
       • the ledger's unsalted original_argv_sha256 makes the <user> redaction reversible (medium, privacy).
   • Open policy question (routed to agent-lab-4d, owner of the verdict pipeline that consumes scripts/platform_status.py): QUALIFYING_STAGES counts install, so a version-only install receipt plus one independent agree can make a Linux winner accepted.
     • Recommendation: accepted needs a use pass, and install-only supports at most conditional.
     • Decide it in its own trust-path PR before Record NativeStack workstation acceptance and retained failures #160 merges; if contested, it goes to the user.
     • Require a reviewed use-stage pass for an accepted platform status #164 (agent-lab-4d): accepted needs a reviewed use pass. A demotion regression reproduced by github-automation-gap-closure is fixed and verified at 76d8282 (a Linux winner accepted through registered evidence stays accepted). agent-lab-e9 is reviewing the rest. It doesn't merge inside the release window.
   • Bind the no-git archive check to the release; report new-machine content drift in release_due; run step 6 under Python 3.13 #163 merged (github-automation-gap-closure; head 94de71f3, merge 81385fe6):
     • Contents: the release-bound archive check; release_due reports content drift over 240 new-machine files (links, linked directories, any top-level path, nested install guides, modes, the manifest's non-pin fields); step 6 under Python 3.13.
     • Reviews: 3 Codex rounds and an Opus lens. All findings are fixed except Extend the pinned-release self-consistency test to release_due's new-machine discovery #168 (extending the pinned-release self-consistency test to this discovery; needs a reference form that is precise at a commit).
     • Re-pin procedure (F3): update the re-pin branch from main right before merging. The ruleset doesn't require up-to-date branches, and the post-merge push run isn't strict. The structural fix is a follow-up.
5. The re-record release (was v2026.09.23.2; the next tag after v2026.09.24.1), owner agent-lab-e9, after Blind lanes round 2: sanitized export, receipt aliases, two-family adjudication #145, the re-record and the ENFORCED_PLATFORMS widening. It carries the finalized 20260923 verdicts and Blind lanes round 2: sanitized export, receipt aliases, two-family adjudication #145's two blind-lane agent definitions, which only lane runs need. It also carries the per-job CPU quota in adoption/tools/ecosystem-bounded-run (Cap each guarded job's CPU where the user manager delegates cpu #158, 7c211add). That is a content change to an existing file, so a host on .1 keeps the old runner until it moves to .2; release_due on main reports current with due: []. Hosts move to it through adoption/update.md "Moving a host to a new release".
   • ETA: Blind lanes round 2: sanitized export, receipt aliases, two-family adjudication #145 about 1 h after a clean review round; the re-record PR about 5-6 h after that.

In parallel

• Record Mac (macos-m5pro-20260924) foundation use receipts: rtk, worktrunk, gitleaks #185: the first real-Mac evidence (macos-m5pro-20260924: use receipts for rtk, worktrunk and gitleaks, plus the hardware profile). Merged as 3699012; its six receipts are accepted after agent-lab-e9's independent review.

  • The workflow runs one Opus judge per receipt, a skeptic, a consistency lens that simulates the derived macOS status after an agree, and a privacy lens.
  • An independent agree would be the first input that moves a macos-arm64 winner towards accepted, under Require a reviewed use-stage pass for an accepted platform status #164's use-stage rule.
  • The result goes on the PR. If it holds, agent-lab-e9 records the host_receipts.py review entries as agreed.
  • Record Mac (macos-m5pro-20260924) application and data use receipts and the macOS pnpm 12.6.0 re-qualification #201 (the Mac's application and data receipts) merged as 67666d6e (2026-09-24T17:32Z). Its only agrees came from reproducing the commands. agent-lab-e9's independent review (wf_f7021e8d-fd3, Record Mac (macos-m5pro-20260924) application and data use receipts and the macOS pnpm 12.6.0 re-qualification #201 comment 5818957285) found blocking issues in all seven receipts. Record independent needs_changes reviews on #201's Mac application receipts #209 records those needs_changes reviews; on macOS it moves fastapi, nextjs, postgresql and playwright-test from host_verified to untested (dissented). github-automation-gap-closure checked the generated diff: exactly those four columns change. State the use-review bar (layer role, positive control, backed claims, bound pin) in the evidence flow's step 8 #210 (github-automation-gap-closure) writes the Record Mac (macos-m5pro-20260924) foundation use receipts: rtk, worktrunk, gitleaks #185 bar into contributing-evidence.md §3 step 8: layer role, positive control, backed claims and a bound pin; reproduce-only is not an agree. agent-lab-e9 reviews it.

• Roster, trading harness (recorded 12:44Z):

  • Paper-only execution-capacity harness for Alpaca order actions (rate-limit-adaptive, websocket state) #165 (order-throughput harness, dd8d3f88) and Mover v3 research plan and protocol draft from the verified 2026-09-24 SOTA sweep (not frozen) #167 (mover v3 plan) are merged code without a live owner; no local session claims them. Change them only through a new reviewed PR.
  • sota-workflow-resolution owns Mover studies: preregistered v1 and v2 (no pass in any family), data corrections D1-D6 #162 (mover studies) and Mover trial mode for the adaptive NautilusTrader paper engine (SYN-verified; paper needs an isolated account) #172 (engine mover mode), and is working on claude/live-capacity-mode-20260924: a separately gated live mode in order-throughput/capacity.py.
    • Its gates: a user-written go file, the user's live key and an explicit flag; RTH only; non-marketable probes; strict caps; stop on any fill without selling.
    • No live run happens without the user's explicit go.
    • Its review is two Opus lenses; Codex is unavailable until 2026-09-29.

• Closed by agent-lab-87 (items from its readiness critic):

  • Scorecard alerts ci: bump actions/setup-python from 6.3.0 to 7.0.0 #29 and ci: bump actions/setup-go from 6.4.0 to 7.0.0 #30 are dismissed as false positives with a written reason. run_g9.sh pipes loopback /v1/models JSON into a fixed python3 -c parser, and the file is hash-registered evidence. 8 Scorecard alerts remain open.
  • The current native-offhost-app-state and native-offhost-restore definitions (setup-python 7.0.0, setup-go 7.0.0) now run green on main 4a4c8a28: runs 35964963227 and 35964966279.

• Still unowned:

  • The settings template registers a context-mode-cache-heal.mjs SessionStart hook that no catalog step installs; context-mode writes it on first boot. Found by source reading only.

• agent-lab .mcp.json portability: agent-lab Reconcile foundation catalogs with current upstream and Claude evidence #49 merged as 7917b40d (github-automation-gap-closure).

  • serena is removed from the project config; the catalog's user-scope profile provides it.
  • SocratiCode's paths use ${HOME}.
  • tools/mcp-conformance/check.sh expands only ${HOME} in stdio command and args. An independent review had found that a general expansion could send the caller's credentials to a server a config names; that is fixed and re-verified with fake-credential probes. --expand-config feeds the e2e v3 mcp-inspector probe.
  • After pulling agent-lab, approve the changed project MCP servers if Claude Code prompts.
  • SocratiCode's host-layout prerequisites are in docs/architecture.md.

• github-automation-gap-closure: Add a 32-layer landscape-sweep lane to manifest-20260923 and source reviews for its newcomers #153 merged (9eac1f97, 2026-09-24 01:18Z, inside the cut-off): 32-layer landscape sweep plus source reviews in manifest-20260923.

  • Sweep. 107 proposals, including the completeness critic's follow-up round.
    • 11 survive both refuters.
      • targeted_candidate: marker, MinerU, claude-code-action, tempo, LightGBM, RD-Agent, trufflehog, ollama.
      • keep_but_compare: PaddleOCR, betterleaks, exa-mcp-server.
    • 96 are refuted, each with its votes.
    • 12 are known to the catalog, including 5 recorded in automation.json.
  • Source reviews. 62 neutral source_review files at pinned commits (upstream words only), registered and listed first in evidence: 11 for the survivors and 51 for the existing non-refuted newcomers, including 6 HF models.
  • New open gaps.
    • osv-scanner and Scorecard: automation this repository runs but its layer rows omit.
    • poppler's catalog URL is a third-party mirror.
  • Reproducibility. main's manifest reproduces byte for byte from the recorded lane inputs (lanes-final-20260923.json) with the unchanged generator.
  • Review.
    • Three independent review rounds, with every medium and low finding closed.
    • Codex's 6 threads (P1/P2) fixed and resolved:
      • RD-Agent's held-out split;
      • TruffleHog's live-verification arm;
      • MinerU's stale figures;
      • Marker's model-weight licence;
      • retaining the stopped run.
    • The final review also corrected PaddleOCR's false rank-1 claim and claude-code-action's prior-record disclosure.
  • Usage.
    • Completed run: 119 agents, 12,123,864 subagent tokens, 2,823 tool uses, 6,615 s.
    • The stopped first run is retained (9 returns; usage unknown).
    • A CI gitleaks false positive (existing pin commit ids once the lane text named Sourcegraph) is closed with a rule-scoped, exact-file allowlist and regression tests.

• Codex restored (about 13:05Z; recorded 13:03Z): the user signed in with a new ChatGPT account through native device auth. codex login status reports ChatGPT, and a test codex exec returned OK (agent-lab-e9). Codex reviews and the re-record's Codex lane are unblocked. Between 10:58Z and 13:05Z the shared account was at its usage limit; Codex steps in that window were unavailable, not skipped.

• Order to the re-record and v2026.09.23.2 (updated 2026-09-24 ~03:15Z). Every step waits on Blind lanes round 2: sanitized export, receipt aliases, two-family adjudication #145.

  1. Add a 32-layer landscape-sweep lane to manifest-20260923 and source reviews for its newcomers #153 merged 9eac1f97. Record the landscape sweep's measured usage (stopped run no longer 'unknown'; completed run's effort verified) #154 (usage evidence; lane_limits only, not in packets) is on auto-merge.
  2. agent-lab Resolve native dashboard, memory lifecycle and evaluation gaps #43 merged 6beff68: every lane stage runs at effort max.
  3. Blind lanes round 2: sanitized export, receipt aliases, two-family adjudication #145 MERGED as ff2ac50a (16:28:15Z; recorded 16:28Z), after 13 independent review rounds; see the history below. A paginated re-count before the merge showed 101 threads, 0 unresolved, and state CLEAN. Low follow-ups are tracked in Blind-lane audit and registry lows left after #145 rounds 9–12 #189, Extend the pinned-release self-consistency test to release_due's new-machine discovery #168 and Enforce lane-provenance.json append-only against the base branch in CI #175.
     • Rounds 1–3 (6c22858a → 3d0943cc) and the delta to 7b717864 are closed, apart from items carried forward below.
     • Round 4 at 2a2ffb8d (four lenses, 18 agents): 13 findings confirmed (2 high, 11 medium), 1 downgraded to low. Details in the scratchpad review145-r4.md.
       • High:
         • F1: id-keyed lists in adoption/manifest.json (profiles' component_ids, recipe_map, "Selected …" labels) and runtime-target.json still single out winners in 7 layers. The new checker reads only URL lists.
         • F2: user-level Agent Skills reached blind Codex children through the inherited HOME. Addressed in a516c477.
       • Medium:
         • Exposure disclosure understated (F3).
         • 5 packet-cited evidence files are missing from the export (F4 = R4-REG-2 = OPS-1).
         • Role-key stripping removes exercised evidence (R4-REG-1).
         • Packet membership fields (F5). Addressed in a516c477 with sealed packet keys.
         • isolated_codex_home can delete the credential when the native and run-scoped auth.json are the same file (BIND-R4-1).
         • A stale refuter events file voids a clean judgment and drops its leak (BIND-R4-3).
         • An interrupted run loses its held leaks (BIND-R4-4).
         • Recipe step 7 leaves validate.py failing (OPS-2).
         • RealExportIsolationTests has no remedy after a wave changes a winner (OPS-3).
     • Since then: c7331916 (Codex review of 2a2ffb8d) and a516c477 (F2 and F5; sealed packet keys restored by record_verdicts, adjudicate, landscape.py and the verdict gate; registry pruned to main plus current code).
     • Round 5 at a516c477 (five lenses, 20 agents; details in the scratchpad review145-r5.md):
       • Of the round-4 IDs, F2 and F5 are fixed and 4 are partly fixed. 30 are still open at a516c477, which is expected because the owner's local 52344da8 fixes them.
       • New high: the run-home rebuild rmtrees <work>/codex-home and deletes the whole native Codex home when that path is, or resolves into, the native home (INT-R5-2).
       • New medium:
         • The packet-keys document isn't bound to what the lanes judged (INT-R5-1).
         • The recipe and README omit --keys-out and --packet-keys, so step 1 exits 2.
         • withhold_prose both under-drops (short names) and over-drops (a layer_scope_terms pointer to a field the packets lack).
         • The verdict gate rejects documented "failed" lane outcomes.
         • Blind children inherit the caller's whole environment (transcript pointer, messaging socket).
         • The archived/license gating order.
     • Round 6 at fe5143e2 (the round-4 and round-5 fixes plus the main merge; five lenses, 13 agents; review145-r6.md):
       • Of the re-checks, 41 are fixed, 18 partly and 1 open.
       • Integrity: PASS (the packet-key binding and the evidence.json merge resolution hold). Codex isolation: PASS. Regressions: lows only.
       • 8 confirmed mediums:
         • Export redaction corrupts evidence (B6-4): 30 hash-bound files mismatch and code listings are mangled.
         • about_candidate drops incumbents' negative evidence (B6-5).
         • A per-packet matcher regression (B6-1).
         • Redaction misses context and short-name choice statements (B6-2).
         • Categorical scalars in cards isolate winners (B6-3).
         • Step 7 omits new_host_grand_list.py --write (OPR6-1).
         • Step 7's rehash hides tampering (OPR6-2).
         • The role-label remedy overwrites the wave's packets (OPR6-3).
       • Recommendation to the owner: never rewrite evidence files; export them verbatim or exclude them, and disclose per layer.
     • Round 7 at a4dfd99e (2b290630, the round-6 fixes c7875d94, and a main merge; four targeted lenses, 13 agents; review145-r7.md):
       • The verbatim-export design is verified: 611 of 633 exported files are byte-identical, and the rest differ only by stripped label keys.
       • Re-checks: 18 fixed, 20 partly, 3 changed, 1 open. Full suite 3903 OK on both Pythons.
       • 7 distinct confirmed mediums:
         • An interrupted codex_lane run keeps a flagged layer, and a resume counts it (REG7-1 = ISO-R7-1).
         • The stricter checker fails 96 of 288 winner changes, so a re-record would not pass CI (OPR7-2).
         • The keys rebuild isn't resumable (OPR7-1).
         • Ordinary words become candidate terms, a regression (BL7-3).
         • Status-free limitations are dropped (BL7-4).
         • The per-layer exposure misses tables (BL7-1), and nothing records it (BL7-2).
       • Convergence plan, proposed to the owner:
         • freeze new detection features;
         • fix the 5 correctness items with real-data tests, and make disclosure the recorded mechanism;
         • run round 8 on the delta only;
         • recommend merge if it finds no new medium, with the residuals disclosed and tracked in follow-ups.
     • Round 8 at 68e74f2c (the convergence round, delta-only; four lenses, 10 agents; review145-r8.md):
       • Re-checks: 23 fixed, 8 partly, 0 open. The disclosure lens passes. Full suite 4078 OK on both Pythons; hosted CI green.
       • 5 confirmed in-delta mediums. One (the prose exposure is optional) is already fixed in the owner's local edb59b5a. The other four:
         • the audit's shell_active_text hides reads that the previous version flagged (REG8-1);
         • an interrupted rerun keeps a flagged return (NEW-1);
         • CI's real-export test fails on report-only hit kinds (OPR8-1);
         • step-7 remedies edit trust files, so a wave PR fails the gate; the fix is a documented two-PR remedy flow (OPR8-2).
       • Round 9: targeted at edb59b5a (the new transcript_audit.py, a Codex P1 exception to the freeze) and the four fixes. The bar is unchanged.
     • Round 9 at 014d046f (three targeted lenses, 7 agents; review145-r9.md):
       • Re-checks: 16 fixed, 2 partly. The transcript audit killed 30 of 37 independent mutants. Full suite 4100 OK; hosted validate-macos FAILS.
       • High F1: a tilde-prefixed Glob or Grep path gets past the transcript audit, which records a blindness violation as clean.
       • Medium:
         • REG9-1: a test hard-codes /home, so the macOS suite fails.
         • BR9-1: the transcript audit is neither a trust path nor bound in provenance.
         • REG9-2: one retried agent voids the lane. Already fixed locally (738fce95 plus per-item retry accounting).
       • Round 10: targeted at those fixes only.
     • Round 10 at 821c23cc (two lenses, 4 agents; review145-r10.md):
       • Fixed and verified: BR9-1, REG9-1 (hosted CI fully green, validate-macos included) and REG9-2. Full suite 4102 OK.
       • 1 high, found by both lenses: whitespace escapes the F1 guard. Claude Code trims Read/Glob/Grep paths before resolving them, and the audit does not. A verified candidate fix flags any value JavaScript trim() would change.
       • Lows are in the file, including a registry CI guard for transcript_audit_py_sha256 and a lane-return schema check (BR10-4).
       • Round 11: targeted at the trim fix, its tests, the registry guard and the BR10-4 check.
     • Round 11 at ca89c0d8 (two lenses; review145-r11.md):
       • The registry lens passes. The trim fix holds for whole values: 0 bypasses, and 0 false positives over 28,601 real values. Full suite 4107 OK.
       • 1 medium (NORM11-1): an absolute Glob pattern whose base ends in .. plus trim whitespace still lists names one level above an allowed root. The verified fix flags any path component trim() would change.
       • The local Codex review's single P2 (registry entries edited instead of appended) was declined with evidence: the entries are branch-only and nothing references the old hash. CI enforcement of append-only is tracked as Enforce lane-provenance.json append-only against the base branch in CI #175.
       • Round 12: a small targeted check of the NORM11-1 rule. If it is clean, merge is recommended.
     • Round 12 at 8007a1a2: PASS (review145-r12.md), with no medium or high finding in the delta.
       • NORM11-1 is fixed: 0 bypasses across 54,508 derived runtime cases and 27 end-to-end forms, apart from the pre-existing low NORM11-3(b). 0 false positives over 28,697 real values.
       • NORM11-2, RR11-1 and RR11-3 are fixed and mutation-checked.
       • Full suite 4111 OK on 3.12 and 3.13. Hosted CI green (19 SUCCESS). The local Codex review of the delta is clean.
       • Recommendation sent to the owner: merge. Merge main first and re-run CI, because main moved to 5af35816 and Blind lanes round 2: sanitized export, receipt aliases, two-family adjudication #145 touches the gate's trust paths.
       • Low follow-ups: R12-1, R12-2/RR11-2, NORM11-3, BR9-2, LOW-1 and others, plus open Extend the pinned-release self-consistency test to release_due's new-machine discovery #168 and Enforce lane-provenance.json append-only against the base branch in CI #175.
     • The first merge attempt was refused. An unresolved Codex P1 (JSON prose was not measured) sat in thread 101 of 101, and a first:100 GraphQL thread count missed it. Thread counts are paginated from now on.
     • Round 13 at 4ab4dfe2: PASS.
       • The JSON-prose exposure fix is verified, and the owner's numbers reproduce exactly: 19 cited files and 21 of 30 layers; 28 of 30 reachable.
       • Both main merges are verified; evidence.json matches on all 5247 entries.
       • Hosted CI is all green (validate 4297 OK). Full suites 4249 OK on both Pythons. The local Codex review of the fix is clean.
       • Paginated count: 101 threads, 1 unresolved (the fixed P1). Go sent.
       • New low R13-1: the 3 layers added to exposure are false positives, so the disclosure over-reports, which is the safe direction.
         3b. Next, in order (agent-lab-4d; github-automation-gap-closure dry-runs each branch):
     1. Rebase Carry the dated manifest's surviving newcomers into every packet (--manifest-newcomers) #151 onto the new main. Done: Carry the dated manifest's surviving newcomers into every packet (--manifest-newcomers) #151 merged as 149940af (2026-09-24T17:21Z). Only manifests/evidence.json conflicts: take main's and rehash Carry the dated manifest's surviving newcomers into every packet (--manifest-newcomers) #151's 4 files. Then validate, full suites, CI, merge.
        • 2026-09-24: rebased as cf82e689. A Codex P2 there: with newcomers, granite-embedding's "embedding" redacted semantic-rag's requirement. Fixed in ba57f96f (19 capability and format words are now generic name parts).
        • Independent targeted check at ba57f96f: clean. On the real packets (default and newcomers), exactly the two intended texts change, and each affected candidate keeps an owner, unique part or whole-name term. Local lane-packet tests: 113 OK on 3.12 and 3.13.
        • Hosted: the 6 required checks pass; the validate run on merge 0ba415e (into main 2d86d55) passed 4321 tests. Threads: 4 of 4 resolved (paginated).
        • Low, optional: pin the LEAN Alpaca Brokerage model defaults restoration in a test. With newcomers, QuantConnect in that candidate's own card becomes a placeholder (already so at cf82e689).
        • Before the merge, main had moved to 9a442a2e (Reconcile trading records with merged evidence (runtime-target, snapshot gate, dashboard, new-host EdgarTools receipt) #200). agent-lab-4d built the exact merge tree (591b9075, which is the squash's tree): validate.py passed and 4321 tests OK on 3.12. Threads: 4 of 4 resolved (paginated). The optional lows are in Blind-lane audit and registry lows left after #145 rounds 9–12 #189.
     2. The 20260923 re-record wave at the merged main (after Carry the dated manifest's surviving newcomers into every packet (--manifest-newcomers) #151). prepare.sh installs the two vendored roles as recipe step 3, from the clean agent-lab worktree /home/seath/code/agent-lab-rerecord (b31f640, descends from e070125). Then the Claude and Codex lanes, collect, adjudicate, record --write/--check, and the wave PR.
        • Plan (agent-lab-4d, from the merged main, because execution-broker's packet changed): a fresh catalog worktree at origin/main; prepare.sh with wave date 2026-09-24 (manifest-20260923, --manifest-newcomers, --withhold-labels); run-lanes.sh (the Claude lane headless, and the Codex lane at gpt-6-astra/high, --jobs 2); collect with --transcripts; adjudicate; record --write then --check; the wave PR. github-automation-gap-closure dry-runs the wave branch.
        • First 2026-09-24 wave (at 149940af) could not be recorded. The Codex lane's blind audit voided 15 of 32 layers, all on false positives. The headless Claude lane was killed at claude -p's 600 s background ceiling. Fix: Enforce a retrieval-CLI-free PATH for blind Codex children; stop voiding lanes on search terms #206 (a restricted blind-child PATH plus audit precision; the recipe sets CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS=0). Review by github-automation-gap-closure, with Opus/max lenses and an adversarial verifier on every finding:
          • Round 1 at f5755c12 (15 agents): 3 medium and 2 low findings in the delta (file:///, one-sided '/'+x, the no-PATH shell default; the npm launcher, stdout parsing).
          • Round 2 at 882ba4b9 (12 agents): all six fixes verified; only low and nit findings remain.
          • 2d7fd6c0 also fixes the literal joins and the URL false positive in the by-path rule, and adds a measured network-block probe (EPERM from the child).
          • 891ab70f closed every round-2 low and both optional items; github-automation-gap-closure's targeted check was clean.
          • Enforce a retrieval-CLI-free PATH for blind Codex children; stop voiding lanes on search terms #206 merged as b122312d (2026-09-24), with 0 of 0 threads (paginated).
          • Refuse env-style shell launchers; keep unix and local URL schemes checked (#206 review lows) [hold: after the wave] #216 (50ae3acc) holds the last two lows: refuse env-style shell launchers; keep file, local and *unix* schemes checked. It is pre-checked clean.
            It merges only after the wave lands, because record_verdicts' current-code check compares the wave's codex_lane hash.
            AF_UNIX was measured too: connect() from the child gets EPERM, and the control connection is accepted.
        • Wave 20260924 re-record running (agent-lab-84, formerly agent-lab-4d). It was prepared fresh at b122312d: 32 packets, export tree 9e70ff19, 0 gating hits.
          Both lanes started at 20:37Z:
          • the Claude lane headless, with CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS=0;
          • the Codex lane at gpt-6-astra/high, --jobs 2.
            The aborted first attempt is kept as layer-verdicts-20260924*-aborted1. After record, github-automation-gap-closure dry-runs the gate and validators.
          • Quiet window from 21:50Z (2026-09-24) until agent-lab-84 says the Claude lane is done (expected around 01:30Z on 2026-09-25): no large Opus/max workflow fan-outs in any session. All sessions share one account's limits.
            At 00:15Z: 125 agents so far, with 0 API errors. All 32 proposals and all 64 refutes are done. 27 of 32 layers needed a revision, and about 54 rechecks remain. The Codex lane is complete (32 of 32 clean at b122312d).
            The Claude lane (d8854dca / wf_9a22db6d-92e) hit the account session limit at 20:54Z: 41 of 50 agents errored, and all 32 layers came back unknown. It reruns in full after the reset.
            It needs about 100 to 130 Opus-max agents. github-automation-gap-closure's Enforce a retrieval-CLI-free PATH for blind Codex children; stop voiding lanes on search terms #206 reviews (27 agents, about 3.3M tokens, 19:03–20:15Z) had used much of that budget beforehand.
          • Codex lane at b122312d: 30 clean and 2 flagged, both false positives ('://' in ref; an rg fragment /runner.py$). Only those 2 are resumed, with the same code. The two patterns go to Refuse env-style shell launchers; keep unix and local URL schemes checked (#206 review lows) [hold: after the wave] #216.
     3. The ENFORCED_PLATFORMS widening PR.
     4. The re-record release (agent-lab-e9).
  4. The re-vendor: done. The lane content landed with Blind lanes round 2: sanitized export, receipt aliases, two-family adjudication #145: the lane-provenance.json entry with prompt_sha256 fa75f113, test_the_lane_prompt_it_fills_is_the_catalog_copy, and layer-verdict-lane.js and blind-adjudicator at max. claude-code-optimal-defaults' branch claude/max-effort-default-20260923 merged as Run every Claude child at effort max under an Ultracode main loop (installer agents, examples, committed project settings) #196 (2d86d55, 2026-09-24T16:45Z) with no lane files. It moves the installer agents to max and adds the committed ultracode project settings, the examples and the contract suites. Its rebase review (ff2ac50a...cad8d255, one Opus/max reviewer; 4 findings, fixed in 152cfa42) ran before the merge. github-automation-gap-closure's post-merge review of 152cfa42, the settings merge and the squash is clean.
  5. The user-level blind roles (blind-lane-reviewer, blind-adjudicator) are in adoption/agents/claude/ on main, so install_claude_profile.py --only agents from a main clone now installs them (it copies every *.md there). agent-lab-17 reinstalls them from catalog main before the re-record.
  6. Carry the dated manifest's surviving newcomers into every packet (--manifest-newcomers) #151 packets built on main, including Record the landscape sweep's measured usage (stopped run no longer 'unknown'; completed run's effort verified) #154, then the 20260923 re-record (agent-lab-17): Claude lanes at max, Codex lane at high.
  7. The ENFORCED_PLATFORMS widening (agent-lab-17, scripts/landscape.py:51). It lands only after the re-record, because it fails main before then. github-automation-gap-closure dry-runs the gate and the validators.
  8. v2026.09.23.2 (agent-lab-e9). After the cut, github-automation-gap-closure verifies it independently from a fresh clone: attestations, verify-asset, step 0, release_due --strict. agent-lab ci: bump actions/setup-python from 6.3.0 to 7.0.0 #29, the readiness snapshot, is agent-lab-e9's alone and is finished after .2.

• sota-workflow-resolution: Give the catalog-freshness job full history so the release-pin tests resolve #150 fixes the catalog-freshness.yml regression after the re-pin. The shallow freshness clone lacked the release commit, so 4 of 3,501 test_release_pin_contents failures skipped propose (run 35931645459).

  • The fix is fetch-depth: 0. It passed review; the review also suggested a workflow-hardening guard test.
  • This is the only full-suite job with a shallow checkout.

• sota-workflow-resolution (trading lane): IBKR paper orders through NautilusTrader 1.231.0's IB engine, with an after-hours plan #147 (IBKR paper orders plus after-hours) and Record the regular-session no-signal run and the first extended-hours paper run #149 (adaptive-paper evidence) are open, and both include the verdict gate.

  • Book same-account paper fills made outside the durable ledger #136 books external same-account fills into the durable ledger. It is done after the merge plus one reviewed apply on the real ledger (backup, dry run, apply, cash reconciles), with the STOP hold archived under a release record. ETA today.
  • Make the adaptive-paper tests hermetic from the host STOP kill switch #137 merged: the adaptive-paper tests are hermetic from the host STOP file.
  • Live record of paper runs: NautilusTrader JSON log, decision stream, read-only manifest #138 adds a live record of paper runs. ETA today; the release does not wait for it.
  • A clean 1x paper E2E in the original ledger with the frozen config, after Book same-account paper fills made outside the durable ledger #136 applies. ETA before the market close.
  • IBKR paper orders through the NautilusTrader 1.231.0 IB engine. The receipt passed and the PR opens tonight. ibkr-local-acceptance stays not_established because of rc5 #4983.
  • The 2x/4x leverage rungs need a separate dated limits decision (not today).
  • For the release, include Book same-account paper fills made outside the durable ledger #136 and Make the adaptive-paper tests hermetic from the host STOP kill switch #137 if they land in time. Do not wait for Live record of paper runs: NautilusTrader JSON log, decision stream, read-only manifest #138 or the IBKR PR.

• Record scheduled memory learning and daily native lifecycle evidence #116, from a Codex session: its secret-scan blocker is resolved on main by Exempt ai-memory's SHA-256 rejection fingerprints (#116's evidence) from generic-api-key #155 (e2f014fd). The 4 false positives are ai-memory SHA-256 rejection fingerprints, exempted by exact value. The allowlist regression tests now run in CI (19 OK). Its Codex owner or the user only needs to merge origin/main into the branch.

• Codex ecosystem lane, owner agent-lab-87 (2026-09-24). This lives in the codex-ecosystem repository. agent-lab-87 is also running a read-only new-PC readiness check (fresh clones of .1 and main).

  • CPU limit in ecosystem-bounded-run: done. Cap each guarded job's CPU where the user manager delegates cpu #158 was squash-merged as 7c211add (04:11Z) from the reviewed head 7a1ffdc5, with every required check green and the PR CLEAN.
    • Review by github-automation-gap-closure:
      • 16 mutants, all killed;
      • the upstream claim that systemd v252 delegates cpu to [email protected] is verified;
      • the runner on main (cde3374b) is byte-identical to the host copy.
    • Still open: the interactive-contention comparison named in adoption/tools/README.md ("CPU quota").
    • It ships to hosts in .2 (see step 5).
  • Retention for per-wave caches: a host-only, report-only ecosystem-wave-retention rule that deletes nothing. It is not repository content: .2 carries only its dated status note in docs/next-host-stages.md. Round 2's one high was that its citation gate read a stale catalog checkout; agent-lab-87 is fixing it on the host, and its round-3 review is running.

Done 2026-09-22/23 (for reference)

• agent-lab-fe (agent-lab-e7): catalog Generate the new-host grand list and record next host stages with upgrade evidence #102, Record the local memory-embedder evidence (benchmark, hybrid re-measurement, live deployment) #127, Keep the grand catalogs current as new hosts and runtimes record evidence #130 and Record this host's vLLM receipt with its qualified local models #131, and agent-lab Publish verified native foundation and scoped OmniRoute installation #15, Document verified Desktop token workflow and reusable task prompt #18 and Save verified native quality defaults and acceptance practice #35. Its exit criterion is met, and it has no open PRs. External: feat(embedding): query/document prefixes for OpenAI-compatible embedders akitaonrails/ai-memory#859, recorded as a pin deviation; it does not block the release.
• Catalog: macOS arm64 clean install: Homebrew prerequisites, darwin pins, launchd agents (brew-services semantics), embedding acceptance, recording-tooling CI #94, Lift the MLX smoke lock past its transformers and mlx advisories #99, Generate the new-host grand list and record next host stages with upgrade evidence #102, Close spy-lean-parity and dividend-sim-module on the qualifying parity v2 replay (sim rung ready) #103, Resolve the first CodeQL default-setup alerts (href allowlist, HTML tokenizer, exact hostname) #104, Close the catalog's GitHub automation gaps: OSV/zizmor scans, gates, immutable releases, target ruleset #108, Record the post-merge verification of the GitHub automation closure #112, Derive platform status from host receipts; independent, pin-bound receipt reviews #117, Split the OSV SARIF upload into a tool-free job; fix the blind comparison's confirmed defects and file its evidence #121, Verdict-pipeline integrity before the 20260923 re-record: per-wave CI, unbiased packets, two-family rule, survivorship, receipt-backed status #122, Fix three engine defects exposed by live SIP paper trials; first passed 1x trial recorded #123, Verdict integrity round 2: bind rows to sealed waves, strict single-lane authorization, immutable waves, sealed refutations, retained packets #124, Make the full test suite portable to macOS (symlinked temp paths, platform-specific tests) #126, Record the local memory-embedder evidence (benchmark, hybrid re-measurement, live deployment) #127, Keep the grand catalogs current as new hosts and runtimes record evidence #130, Record this host's vLLM receipt with its qualified local models #131, Integrate gap wave 2 (agent-lab-17): 23 layers, 189 receipts, owner ledger #132 (gap wave 2), Fail the Codex lane with a setup message when the codex CLI is missing #133, Make the new-PC install guides followable from the pinned release; add the release-move flow, a docs-consistency test and a receipt-staleness report #134 (new-PC guides, release-move flow, docs-consistency test, receipt-staleness report), Add the required verdict-review-gate: sealed cross-family evidence for every changed verdict row #135 (verdict-review-gate), Make the adaptive-paper tests hermetic from the host STOP kill switch #137 (hermetic STOP tests), New-PC Claude Code user profile: native claude-code pin, guard, agents, user-scope MCP, settings apply #139 (new-PC Claude Code user profile, cd5db091: native claude-code 2.1.280 pins, adoption/{hooks,agents,mcp}/claude, install_claude_profile.py, --configure-claude-user-profile; closes the Open native terminal profiles and map all foundation installation layers #34 gap), Blind both layer-verdict lanes for the 2026-09-23 re-record #141 (blind lanes), Let lane packets carry each component's registered receipts #142 (lane packets carry registered receipts) and Make verdict-review-gate apply the landscape validator's withheld-key policy #144 (one withheld-key policy for the gate).
• agent-lab: Publish verified native foundation and scoped OmniRoute installation #15, Document verified Desktop token workflow and reusable task prompt #18, ci: maintain native research automation and attest catalog archives #26, Fix guest reboot observation and preserve failed native evidence #27, Accept scoped native recovery and remove repeated paper approval #32, Open native terminal profiles and map all foundation installation layers #34, Save verified native quality defaults and acceptance practice #35 and Make native dashboards show useful scoped data and command evidence #37 (the two evaluation-fixture Dependabot alerts dismissed as tolerable_risk, time-bounded by the OSV ignoreUntil 2026-10-22; 0 open).

User-only items

• The first physical Mac run.
• Revoking the leftover claude-code-action setup token.
• Commit signing, only if wanted; the required_signatures rule stays off until then.
• The Databento key.
• The live-go decision.
• IBKR steps 2-4, blocked upstream by nautilus#4983 and rust-ibapi#764.

Known local-only noise

The 32 adaptive-paper test failures on the authoring host came from the engaged live paper STOP file. #137 (merged) fixes this. Hosted CI does not have the file.