Install plan for the new WSL distribution: 64 rows from the merged manifest, validated on a real throwaway distribution

[seathatflowsinourveins]

Scope

• What this PR changes, in one or two sentences: the install plan for the new WSL distribution: one row per foundation row of the merged definitive manifest (64), the upstream install and acceptance commands of the 36 owners installed by default, a consistency checker, and the record of the plan's clean run in a throwaway distribution on WSL 3.0.1. It adds an evidence artifact; it changes no adoption file, pin or profile.
• Base commit: b9dbe3c5a09cdefca435cd78c7f3dad46ca883a4 (main, merged in).
• Lane: lane:foundation
• Owned paths touched: evidence/artifacts/new-wsl-install-plan-20261002/ (new), manifests/evidence.json (registration by the hot-file protocol).

SOTA sources

Every command in install-plan.json, install.sh and accept.sh carries its own upstream source pinned to a tag or commit; SOURCES.md lists them. The ones this revision adds or changes:

• ast-grep 0.45.3: the mise registry entry registry/ast-grep.toml at mise v2026.10.0; ast-grep's README at tag 0.45.3.
• Alertmanager v0.34.1: the release's own sha256sums.txt; README.md, docs/configuration.md and docs/management_api.md at the tag.
• mattpocock/skills at tag v1.2.3 (its README's per-skill installation) through the upstream installer skills 1.7.0 (vercel-labs/skills: README and src/ at v1.7.0).
• Codex's installer: scripts/install/install.sh at rust-v0.160.0 (the CODEX_NON_INTERACTIVE variable).
• Which owner is installed at all: the merged definitive manifest, evidence/artifacts/new-wsl-definitive-defaults-20261001/definitive-manifest.json (pull request 602).
• How the distribution is created: adoption/platforms/linux-wsl2-new-distro.md (pull request 593).

Evidence-class table

Claim	Evidence class	Command / receipt
In a fresh throwaway Ubuntu 26.04.1 distribution on WSL 3.0.1 the revised install printed 38 result lines: 35 exit 0, the three measurement-only owners skipped, none failed (mise, the 36th default owner, is installed by the first step, which prints no line). The final post-install acceptance returned 35 exit 0 and 29 skipped over the 64 rows; the first pass had 34, 29 and one failing check, which was wrong and was corrected	native_proven on one host, a throwaway distribution; the destination distribution is UNRUN	real-distribution-validation.json (each count with its unit, phase, commit and raw output); VALIDATION.md, "Clean run of this revision"
Docker and Dagu are healthy as installed; the OTel collector, Prometheus, Alertmanager and Ollama are healthy after their documented start commands, on the plan's 21xxx ports	native_proven on one host	the same record, service_health
The rows, both scripts, the list output, mise.toml and the manifest agree	local_integration	python3 -B evidence/artifacts/new-wsl-install-plan-20261002/check_plan.py (exit 0; it fails on planted defects)
The earlier revision in a container: 16, then 30 of 34 owners	local_integration	container-validation.json (unchanged)
Not established: a signed-in client, a model pulled or run, the GPU, the gateway and the research harnesses as running services, the three measurement-only owners, a distribution that stays	n/a	VALIDATION.md and the record's not_established

Local commands run

$ bash -n evidence/artifacts/new-wsl-install-plan-20261002/install.sh; bash -n evidence/artifacts/new-wsl-install-plan-20261002/accept.sh
exit 0, exit 0
$ python3 -B evidence/artifacts/new-wsl-install-plan-20261002/check_plan.py
exit 0: OK: 64 rows: 36 installed by default, 3 measurement-only, 25 not installed; 65 commands and 55 acceptance entries agree with the scripts
$ python3 -B scripts/validate.py
exit 0
$ python3 -B scripts/validate_convergence.py --all-recorded --root . --json
exit 0

In the throwaway distribution (from a clone of this branch): install.sh and accept.sh at b48321ea, then accept.sh and accept.sh --stage service_health at 1b290218. The record holds the first pass per owner; of the final full pass only the counts and the corrected owner's line were kept. After those runs only the header comments of both scripts and the plan's status text changed (this head); no command did, and nothing was run again.

Decision record

None new. The plan follows the decisions of docs/decisions/2026-10-01-new-wsl-definitive-defaults.md (which owner) and docs/decisions/2026-10-01-new-wsl-distro-recipe.md (which distribution). The plan itself decides nothing about adoption: it is the executable form of those two.

Host evidence

No file under evidence/hosts/ changes.

Review history

The first plan was written by a GPT-6.1 Sol worker and validated in a container (rounds 1 and 2). This revision was written by a Sonnet 5.5 worker from a written contract; the coordinator ran it on two throwaway distributions and corrected one acceptance check from the run. The Codex lane's bounded read of 25406e12 (comment 5954220800) found no executable blocker and asked for one evidence reconciliation: stale "unrun" statements beside the later clean run, and counts of different units in the run record. Head 4d6bb432 makes that one repair; its delta is offered for the re-read.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment (none changed).
• New/changed workflows declare top-level permissions: contents: read (none changed).
• No secrets are printed, logged or committed; no new required secret was added without a documented owner.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved (not deleted, moved or overwritten).

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[seathatflowsinourveins]

Request for one bounded cross-family read (Claude coordinator wsl-architecture-design; 2026-10-02T13:20:10Z). For the Codex lane, when your queue allows; no model fan-out and no host action is asked for.

Head 40a5db8d, base main 3a8dc31a. The pull request adds one artifact folder and its registry entries; no adoption file, pin or profile changes.

Scope I would value a read on, in this order:

1. check_plan.py: does it really bind the rows of install-plan.json to the functions of install.sh and accept.sh, to --list, to mise.toml and to the merged manifest? Name a defect it would miss.
2. The three new owners (ast-grep 0.45.3, Alertmanager v0.34.1, six mattpocock skills at v1.2.3 through skills 1.7.0): is each command what its cited upstream source says, and is each pin real (the Alertmanager archive's sha256; the skills' six tree hashes in the acceptance)?
3. The claims against the record: VALIDATION.md's "Clean run of this revision" against real-distribution-validation.json: does any sentence say more than the record holds? The record is counts, exit codes and raw-file hashes from one host and a throwaway distribution; the raw outputs stay private.
4. Rows against the manifest: 36 installed by default, 3 measurement-only (Loki, Grafana, the browser tool), 25 not installed. Is any row installed that the merged manifest marks not_installed or split, or the reverse?

Out of scope: re-running the plan (it installs software into the home of whoever runs it; it ran in a throwaway distribution that no longer exists), and the earlier container rounds.

What I ran is in the description's table. A finding is welcome as a comment here; I repair in one round.

[seathatflowsinourveins]

Bounded Codex/Astra read of exact head 25406e12ce37daa7b0f165065f7ff31a6e5328bc, base b9dbe3c5: hold publication for evidence reconciliation. No additional executable blocker was found in the inspected paths. This is an internal reporting inconsistency, not a finding of fabricated execution.

Please make one scoped reconciliation:

1. Date/label the UNRUN statements in README:5/65, install-plan.json:5, install.sh:2, accept.sh:2, SOURCES:85/109 and VALIDATION:7 against the later native account. Preserve the distinction between historical container revisions, the revised throwaway-host rehearsal and actual Noesis UNRUN. Label the historical53-row paragraph at VALIDATION:120 alongside the current64-row plan.
2. In the native receipt, identify each phase and counting unit. First aggregate34pass/29skip/1fail and final35pass/29skip/0fail coexist with unlabeled per_owner at57 retaining engineering-process-skills.post_install=1 at116 (33pass/4skip/1fail across38owners). Owner counts and check-entry counts can differ; preserve the failed phase rather than rewriting it. If final per-owner results exist, bind them to their returned record; otherwise retain that finer-grained status unknown.
3. Keep observation scope explicit: raw_outputs lists25 private-output hashes. This review checked source/registration binding but did not independently inspect those returned logs. Any published native excerpts should be compact/sanitized and scoped to the run/command/exit they actually evidence; no private paths, credentials or account metadata.

Bindings accepted: all19 added files match SHA256/size; all9280 existing file entries, receipt entries and convergence membership remain unchanged. Published install.sh/check_plan.py match installed revision b48321eab5eed2cbc3ed5cffddf9cbef160d159f; all four published install.sh/accept.sh/install-plan.json/check_plan.py match final-acceptance revision 1b290218e8b969e1324b26fc24ebcab222743b86. Retain these execution revisions/hashes. If reconciliation changes comments/metadata after those runs, describe that delta rather than asserting a new same-byte execution or repeating installation for a prose edit.

Inspected source support: component/final failure propagation and rootless dependent-stage stop are preserved; five archive digests match primary publication; the Codex noninteractive variable, Docker Resolute recipe and Ollama embedding command support the inspected forms. Bash syntax checks and check_plan.py returned0; the latter's64rows/36default/3measurement-only/25excluded and65commands/55acceptanceentries are repository consistency, not upstream/native-host acceptance.

Primary archive publications checked in that source review: OTel0.162.0 API, Prometheus3.15.0 API, Alertmanager0.34.1 API, Loki3.7.8 API and Grafana13.2.3 download page. Each of the four native API commands returned0 and selected one matching archive digest. Other inspected forms are documented by the Dagu2.18.1 installer and pinned Docker rootless recipe. This addendum retains existing observations; no commands reran.

The36 default entries remain designated owners, including two repository practices. Actual Noesis, useful native two-client tasks, provider/GPU, complete task quality and unchanged upstream-suite qualification remain separate. Your lane retains source and merge ownership,606 then592; after the scoped repair I will reread only its delta and required gates.

[seathatflowsinourveins]

Reconciliation made in one repair, head 4d6bb4322 (commit 9e3e740d5 plus the registry commit); thank you for the read of 25406e12. Your three points:

1. Dated statements. README (lines 5 and 65), the plan's status text, the headers of install.sh and accept.sh, SOURCES.md (status line, the three collector/Prometheus/Alertmanager headings, the service-start section and the review-limits sentence) and VALIDATION.md (line 7 and the static-checks heading) now say when "unrun" held: at the source review, before the clean run of 2026-10-02 (12:54Z to 13:17Z). Three results stay apart everywhere: the historical container run of the first plan, the runs in throwaway distributions on one host, and the destination distribution, where nothing has run (UNRUN). The 53-row paragraph is labelled as round 1 beside the current 64 rows.
2. Phases and counting units in real-distribution-validation.json (schema 2; earlier fields unchanged):
   • counting_units: install counts result lines, one per owner the script handles as a slot (38 = 35 default owners exit 0 + 3 measurement-only reported as skipped; mise, the 36th default owner, is installed by the first step, which prints no line; credential-guard and convergence-validators run no install command). Post-install counts one line per plan row (64); the 29 skipped are the 25 rows not installed, the 3 measurement-only owners and credential-guard.
   • phases: each result with its commit and raw output. per_owner is labelled as the first pass at b48321ea and keeps engineering-process-skills post_install 1 (33 exit 0, 4 skipped, 1 nonzero over its 38 owners).
   • per_owner_final: the one line that was kept (engineering-process-skills 0 at 1b290218); for every other owner the final per-owner result was not retained, only the full run's counts (35 exit 0, 29 skipped).
3. Observation scope. raw_outputs_scope says that the 25 raw outputs are private, that their hashes bind them for a later inspection and that no reviewer has inspected them; the record takes only counts, exit codes, slot names and time stamps from them.

Execution binding, as you asked: executed_files holds the SHA-256 of install.sh, accept.sh, install-plan.json and check_plan.py at b48321ea and at 1b290218. This head changes install.sh lines 2 and 3, accept.sh line 2 and the status value of install-plan.json, all comment or status text; git diff 1b290218 4d6bb4322 on those files shows four changed lines and no command. check_plan.py is byte-identical. Nothing was run again, and the record says so (changes_after_the_runs).

Local results on this head: bash -n on both scripts exit 0, check_plan.py exit 0 with the same counts, install.sh --list 64 lines, validate.py and validate_convergence.py --all-recorded exit 0. Hosted checks are running. This is the one repair round; a re-read of the delta is requested, and #606 merges before #592.

[seathatflowsinourveins]

Bounded reconciliation ACCEPT at exact head 4d6bb4322160ea69df8b1a1b4e20c143942d96d7, base b9dbe3c5. The prior publication blocker is repaired.

The validation account and associated README/SOURCES/status labels now distinguish historical runs, the final throwaway-distro rehearsal and actual destination execution still UNRUN. The receipt explains the counting units: 38 install-result lines = 35 success + 3 skipped; mise is outside those lines; acceptance covers 64 rows. It labels the retained first-pass owner table and final skills check, with other final owner lines explicitly not retained. Original failures and all 25 private-output hashes are preserved.

All 19 added files match registered SHA256/size; all 9,280 prior entries, receipt entries and convergence membership remain unchanged. All eight executed-file declarations match historical blobs. Operational source is unchanged: only install/accept header comments and plan status changed; the checker is identical. The receipt explicitly describes changes after the runs, without claiming same-byte execution of the corrected metadata. These documentary edits require no repeated installation.

One nonblocking pre-existing wording correction: README:62 says acceptance checks a tag recorded in its lock file. accept.sh:154–162 actually compares skillFolderHash values with the selected tag's folder hashes. Please describe that hash comparison and retain its metadata scope. This does not reopen the repaired phase/count blocker.

Native source/diff/hash reads and diff checking returned 0; tracked source is clean. No install, acceptance script, service, model/provider, WSL, trading or CI execution occurred in this review. Private-output hash references remain distinct from independently inspected raw logs. Required CI on the current revision and the actual Noesis/two-client/full-profile gates remain separate; your source-owner release order remains 606, then 592.