Integrate gap wave 2 (agent-lab-17): 23 layers, 189 receipts, owner ledger - #132

Summary

This PR integrates gap wave 2 (gap-wave2-20260923): the executed checks for the executable_now gaps that crosswalk #85 (rows at 92bb279) assigned to agent-lab-17. It covers 23 layers in both catalogs. Each layer ran in its own worktree, and an Opus evidence reviewer reviewed its receipts through fix rounds. No verdict, pin or gate changes here; the 20260923 cross-family re-record consumes these receipts.

Owner ledger (catalogs/landscape/gap-wave2-20260923--agent-lab-17.json, with its view in docs/): 189 receipts.

The branch is one squashed commit on current main. Earlier pushes of this branch carried raw session captures. The squash keeps them out of the published history, but GitHub may still serve those old SHAs, so that is reported to the owner.

Review fixes included

Privacy (Codex P1).

• 21 conversation-derived records are retained host-local and owner-only, pinned by sha256 and marked published: false: session tool-call logs, transcript extracts, tool-result dumps, and process listings showing session environment.
• Username, hostname and LAN addresses in raw outputs become <user>, <host> and <lan-ip>.
• Every pin and generator is updated, and the generators rerun byte-identically.

Promotion gate (Codex P2 ×2, independent review).

• DuckDB volume is read as exact text.
• Only a BASE TABLE in the hashed file is accepted, resolved through system.information_schema.
• A .wal is rejected before and after the read.
• The README scopes exactness to volume.

duckdb_gate_check.py exits 1 unless every case matches (Codex P2).

OSV (Codex P2).

• The portfolio-risk reproducibility lock is scanned: 56 packages, no issues.
• Captured probe manifests are *.captured.
• The supply-chain positive control is requirements.txt.fixture, because dependency review parsed its original name as a manifest.

Evidence hygiene.

• Not-JSON captures are .json.txt, for validate_convergence.
• CodeQL findings in blueprint scripts are fixed. The one false positive was dismissed with its reason, because receipts pin the script bytes.

Incidents disclosed in the layer READMEs: the Codex hooks in workers and document-retrieval, the durable-memory store, and scheduling-supervision (decision recorded: removed).

Checks (local, this head 9b5530e)

• python3 -m unittest: 3318 tests OK, 342 skipped. It ran with the promotion-gate venv and REQUIRE_PROMOTION_GATE_VENV=1.
• Every run: step of the CI validate job passes locally except actionlint, which is not installed here. That includes validate_convergence --all-recorded, both gap ledgers, build_verdicts --check, component_matrix --check, new_host_grand_list --check, gap_crosswalk --check, zizmor and the explorer build.
• Guarded gitleaks on origin/main..HEAD (.gitleaks.toml, --max-target-megabytes 2): no leaks.
  • One 2.5 MB numeric candidate ledger is over the bound and was not scanned. A focused rg finds no key, secret or token patterns in it, but coverage there is incomplete.
• An rg over every changed file finds no username, hostname or LAN address.

🤖 Generated with Claude Code