Commit 0d5e650

"scope": "The install manifest for a new WSL2 distribution: one row per layer of the two maintained catalogs (the 20 foundation layers of catalogs/foundation/manifest.json and the 12 us-equities layers of catalogs/landscape/research-state.json) plus cross-cutting rows for the distribution itself, the runtime workers, the GPT-6 harnesses, the credential practice and the convergence practice. Each row carries the source host's selection of record at its pin of record (what the catalogs select and this host's stack pins or installs: a record of the source host, not a merit result), or a layer's verdict selection that the source host never installed, the verdict under the rules below, the row's evidence class (none_recorded when any winner's acceptance is none_recorded; otherwise a class that at least one winner's acceptance carries, and where the winners carry several, the one listed last in the policy table; stronger evidence for one winner stays in the reasons), the reasons with their sources, and the ordered steps a session runs on the new distribution. The new distribution's selection is decided by the repositories' own quality, not by this host's installed state, pins or integration holds: where the verdict is comparison_required the merit winner is undetermined, the row names the arms, each arm installs fresh by its upstream command, and the comparison's result selects. A winner's acceptance class describes a run that the cited source, or one file that source links, shows was run on a host and what it returned; a check that is only prescribed, planned, not run or failed is none_recorded. Every evidence class describes historical evidence from the source host: a new distribution collects its own evidence, and nothing in this edition is acceptance on it. Upstream currency was read on 2026-10-01 with one bounded read per pinned winner (the GitHub releases page, or the named registry page where the project publishes no GitHub release). Paths not yet on main are cited through their pull request and are not hashed. Citation refresh of 2026-10-04: the pin_source locators resolve the current component entries at main 14048b840; this edition retains its dated pins, which may differ from those entries and are displayed as pin drift. Original sources for the changed dated pins: codex: https://github.com/seathatflowsinourveins/native-agent-stack/blob/5597f9fae2c6148037a74b4e73d2c55fc11c9ff6/manifests/stack.json#L297; rtk: https://github.com/seathatflowsinourveins/native-agent-stack/blob/5597f9fae2c6148037a74b4e73d2c55fc11c9ff6/manifests/stack.json#L822; mcporter: https://github.com/seathatflowsinourveins/native-agent-stack/blob/5597f9fae2c6148037a74b4e73d2c55fc11c9ff6/manifests/stack.json#L625; nextjs: https://github.com/seathatflowsinourveins/native-agent-stack/blob/5597f9fae2c6148037a74b4e73d2c55fc11c9ff6/manifests/stack.json#L1715. Readiness publication boundary: The official baseline uses the #700 bar. Captured-status counts and model READY leads remain source_review; they are not measured qualification at that bar. official30/80 remains historical. CC provisional captured-status30/80 asof11:44Z excludes custody13:02Z. It includes difftastic, structural-search, inspect-ai, harbor and codex-sdk at their captured status despite fresh-session use not evidenced; it is not measured qualification on the #700 bar. Earlier37/40/44/32proposals are superseded by review-715-corrections.json#/bounded_repair_rulings; organic native-arm count requires final verified E2E S4. See docs/decisions/2026-10-05-ns2604-foundation-requalification.md.",

"notes": "Harbor's prerequisites (rootless Docker, the egress allowlist, the Node v22.23.3 pre-step and the native OAuth token) come from the Gate A harness lane; the subordinate-id observation is in the Harbor receipt of PR #570, and the rest is not yet in the repository, so the pilot's record should land it. Inspect AI is the fallback runner. Harbor is the runner because top rule (e) requires an upstream harness and it is the only one run end to end on this stack, not because it beat Inspect AI, which was never compared; Inspect AI stays an arm of any comparison. Wave-5 destination owner cells have no recorded destination acceptance; original host/reference winner evidence remains scoped to those winners. Dated NativeStack2604 documentary re-qualification: official historical30/80 remains until the command center final verified same80-slot E2E; current/conditional values are null and earlier37/40/44/32proposals are superseded by review-715-corrections.json#/bounded_repair_rulings; CC captured-status provisional30/80 asof11:44Z excludes custody13:02Z. It includes difftastic, structural-search, inspect-ai, harbor and codex-sdk at their captured status despite fresh-session use not evidenced; it is not measured qualification on the #700 bar. The official baseline uses the #700 bar. Captured-status counts and model READY leads remain source_review; they are not measured qualification at that bar. no organic native-arm count exists until final verified E2E S4. Sources: evidence/receipts/ns2604-requalification-20261005.json and docs/decisions/2026-10-05-ns2604-foundation-requalification.md. This source projection runs no native destination acceptance and preserves main owner/winner scopes."

"notes": "Install the collector examples under observability/collector only if the OTel route is chosen for per-subagent evidence. Dated host source observations are in evidence/receipts/ns2604-roadmap-evidence-20261005.json: user Alerting attestation stands; initial skip, modified delivery/cache, fresh unmodified defect14 boundary and session idle-period/ccusage-control gaps remain separate historical records. Main plan repairs or pin changes do not qualify them by receipt replay."