Follow-up: anti-pattern log rows for 2026-10-02, install plan README wording, new WSL handbook in the guide navigation

adoption/README.md

@@ -37,7 +37,7 @@ which you install through their recipes ([bootstrap step 2](bootstrap.md)).
 | Adoption profile | Selects | Linux pins | macOS pins | Next native acceptance |
 | --- | --- | --- | --- | --- |
 | `foundation-cpu` | Codex, Claude Code, Context Mode, RTK, QMD BM25, explicitly scoped ai-memory, MCPorter | all 7 | all 7 (5 of 7 at `v2026.09.26`) | Native client setup; one useful context/document call and scoped memory retrieval. On macOS use `macos-arm64-foundation` |
-| `new-wsl-clean-foundation` | [Clean replacement WSL source profile](new-wsl-profile.md): Codex 0.159.3 and Claude Code 2.1.284, with the existing bootstrap's pinned Node/uv/gh prerequisites and CPython 3.13.15 via uv | all 2 | all 2 | Source-review recommendations only; optional rows and comparison arms are excluded from the default install. Pin coverage counts do not qualify either host. New-host acceptance and the dual-image recipe remain pending |
+| `new-wsl-clean-foundation` | [Clean replacement WSL source profile](new-wsl-profile.md), with the generated [new WSL handbook](../docs/new-wsl-handbook.md): Codex 0.159.3 and Claude Code 2.1.284, with the existing bootstrap's pinned Node/uv/gh prerequisites and CPython 3.13.15 via uv | all 2 | all 2 | Source-review recommendations only; optional rows and comparison arms are excluded from the default install. Pin coverage counts do not qualify either host. New-host acceptance and the dual-image recipe remain pending |
 | `macos-arm64-foundation` | macOS only, drafted: Codex, Claude Code, Context Mode, ai-memory, MCPorter, llama.cpp Metal embedding, Qdrant, SocratiCode | 6 of 8 (5 of 8 at `v2026.09.25.2`) | all 8 | The macOS acceptance lane on [the macOS page](platforms/macos-arm64.md); not accepted |
 | `token-efficiency` | Accepted 2026-09-30 as the selection ([routing record](../docs/decisions/2026-09-30-task-model-routing.md)), and changed after `v2026.09.25.1`, whose manifest lacks it: the selected token practice (RTK, Context Mode, explicit-file Repomix, guarded Headroom and TOON, ccusage, QMD, MarkItDown, Serena, SocratiCode, ai-memory, MCPorter) wired into both native clients | all 14 (8 of 14 at `v2026.09.25.2`) | all 14 (6 of 14 at `v2026.09.26`) | The [coverage check](../docs/token-efficiency-stack.md#coverage-check) for command presence and client wiring, then pinned-version evidence and one useful native call per tool in each client, recorded through a PR. Both bootstraps now pin all of it; the macOS pins changed after `v2026.09.26` |
 | `research-runtime` | Historical hash-locked SDK/DuckDB, Dagu and LEAN comparison lane | 2 of 11 | 2 of 11 | Reproduce the retained comparison; this profile does not override the Nautilus destination |

catalogs/landscape/manifest.json

@@ -83,6 +83,10 @@
     {
       "path": "blueprints/catalog-clean-install/README.md",
       "label": "Clean install checks"
+    },
+    {
+      "path": "docs/new-wsl-handbook.md",
+      "label": "New WSL handbook: source recommendations and acceptance stages"
     }
   ]
 }

docs/harness-defaults.md

@@ -215,3 +215,11 @@ When a claim or action proves wrong, correct it where it was relayed and record
 | 2026-10-01 | Citing a second-party reproduction that was a quotation | The coordinator told the user that semble's benchmark score had been reproduced by a second party (LightOn, 0.853 against 0.854). The decision round's critic read the source: LightOn's pull request says the semble and BM25 numbers "come from semble's README" and were not re-run, and the quoted row is an earlier model. | Before calling a number reproduced, read how the second party obtained it: a re-run, or a quotation. Two tables that agree to three digits are a sign of copying, not of replication. | `evidence/artifacts/new-wsl-definitive-defaults-20261001/foundation-definitive.compact.json` (code-search, corrections_by_critics) |
 | 2026-10-01 | Reconciling a second model family that ran with fewer tools | The first blind GPT-6.1 Sol judge of the memory layer ran without shell network or page fetching (`gh auth status` exit 1, `curl` exit 6) while the Claude judges had both. It picked a repository the Claude judge had left out for facts it could not see (created a month earlier, no release); the rerun with network access dropped that pick. | Before reconciling two families, read the second family's event log for failed tool calls and its own evidence gaps, and rerun a degraded unit as a separately labeled unit, keeping the first. Identical packets with unequal tools produce disagreements that come from missing evidence. | `evidence/artifacts/new-wsl-definitive-defaults-20261001/gpt-memory-first-round-units.json`; `docs/decisions/2026-10-01-new-wsl-definitive-defaults.md` |
 | 2026-10-01 | Copying a reviewer's own evidence-gap text into a blind packet | The decision round's memory packet gave the deciders the second first-round reviewer's evidence gaps verbatim. Two of those sentences carried routing wording (an "Astra/Max arbitration" attempt and an approval-policy note), from which a decider could tell which model family had written the notes. The cross-family read of the pull request found it; the packets' own check searched only for model names. | Strip a reviewer's process wording (model routing, tool approvals, budgets) from anything a blind judge reads, keep only the factual limitation, and test packets for routing terms, not only for names. A frozen packet that was already used is published as it was, with the leak recorded. | `evidence/artifacts/new-wsl-definitive-defaults-20261001/README.md`; `tests/test_new_wsl_definitive_defaults.py` |
+| 2026-10-02 | Keeping owned worktrees and unpushed receipts in a directory that is emptied at boot | A WSL restart emptied the temporary directory that held the coordinator's scratch worktrees, two unpushed model-server gate folders and three running reviews. Both gate folders had to be rebuilt from retained raw runs, and the original confirmatory receipt is lost. | Put an owned worktree in a durable directory outside every checkout, push a worker's evidence to a branch when the worker returns, and keep only disposable files in a scratch directory. | this log; `docs/decisions/2026-10-01-new-wsl-definitive-defaults.md` records the rebuild and the loss |
+| 2026-10-02 | Embedding double quotes in an argument that Windows PowerShell hands to a native program | The new-distribution recipe's F10 loop and an R1 probe carried a double quote inside the single-quoted argument of `wsl.exe`. Windows PowerShell 5.1 does not escape it for a native program, so the argument broke apart in the rehearsal (`unexpected EOF`, `stat: missing operand`). The commands had been reviewed as text and had not been run from PowerShell. | Write a command that crosses PowerShell into `wsl.exe` without an embedded double quote, and run each recipe command once from the shell the page names before the page merges. | `tests/test_wsl_new_distro_recipe.py` (`powershell_quote_errors`, with a control that restores the quoted form); `evidence/artifacts/new-wsl-rehearsal-20261002/runs-on-wsl-3.0.1.json` |
+| 2026-10-02 | Counting log lines by a pattern that also matches the system's own configuration echo | The recipe's storage-error count matched `hv_storvsc` and left out the driver's registration line, but not the kernel's command-line echoes, which name `hv_storvsc.storvsc_max_hw_queues=4`. On kernel 6.18.40.1 every boot logs them, so a healthy boot counted as storage errors. | Before a count becomes a stop rule, read the lines it matches on a healthy run of the release in use and exclude the configuration and registration lines. | `tests/test_wsl_new_distro_recipe.py` (`test_all_three_storage_filters_exclude_the_command_line_echo`) |
+| 2026-10-02 | Stating a recurrence that the observations do not cover | The recipe page and its decision record said that `systemd-binfmt.service` fails at every boot of the new distribution. The failure had been observed on one host, and WSL's source shows that it depends on the binfmt protection, a default-enabled and best-effort lock, so it is conditional. The cross-family read of pull request 593 found the overstatement. | State on which host and under which setting a behavior was observed, and keep "every" and "always" for a mechanism that upstream source shows to be unconditional or for a measured series. | this log; the qualified wording in `adoption/platforms/linux-wsl2-new-distro.md` and `docs/decisions/2026-10-01-new-wsl-distro-recipe.md` |
+| 2026-10-02 | Editing a convergence record's frozen input in a pull request that does not own the record | A wording fix to the new-distribution recipe page was made in the terminal-profile pull request (604). The page is a frozen input of the recipe's convergence record, so the hosted validate job failed with `frozen_inputs.sources[0]: SHA-256 mismatch`. The local `scripts/validate.py` had passed, because it does not read the records' input hashes. | Change a frozen input only in the pull request that owns its record and refreezes it, and run `python3 scripts/validate_convergence.py --all-recorded --root . --json` beside `scripts/validate.py` before a push. | `.github/workflows/validate.yml` (the step "Validate declared convergence evidence and scope" failed on this mistake) |
+| 2026-10-02 | Writing an acceptance check against an optional field without seeing the tool's output on a real install | The install plan's check of the engineering-process skills compared a `ref` field of the installer's lock file with the pinned tag. The check had not run before the first install on a real distribution. The lock that this run returned had no `ref`, so the check failed; the installer at the pinned release supports that field as an optional branch or tag, and why it was absent in that run is unresolved. | Take an acceptance check's fields from the tool's own output on a real install, and count the check as validated only after it has run there. The check now compares the lock's recorded `skillFolderHash` values with the expected tree hashes of the skill folders at the tag; it does not hash the installed directories. | this log; `evidence/artifacts/new-wsl-install-plan-20261002/VALIDATION.md` and `real-distribution-validation.json` record the failed check and the later result |
+| 2026-10-02 | Running an upstream installer unattended without its non-interactive switch | The install plan piped Codex's installer to a shell. The installer asks whether to start Codex on the terminal device, not on standard input, so the run on a real distribution waited at the prompt until it was stopped; the earlier container run had no terminal and had not shown it. | Read the installer's source at the pinned tag for prompts and set its documented non-interactive variable (`CODEX_NON_INTERACTIVE=1` in `scripts/install/install.sh` at `rust-v0.160.0`); validate an unattended install where a terminal device exists. | this log; `evidence/artifacts/new-wsl-install-plan-20261002/install.sh` sets the variable and `VALIDATION.md` records the run |
+| 2026-10-02 | Pinning with a reference syntax the client does not have | The install plan added the skills marketplace as `owner/repository@<commit>`. The client reads the text after `@` as a branch or tag, and the shorthand cloned over SSH, which failed on the host key in a fresh distribution. | Check a pin's syntax in the client's own help or source before writing it; where the client cannot pin a commit, add the source by HTTPS URL and compare the cloned HEAD with the reviewed commit in the acceptance check. | this log; `evidence/artifacts/new-wsl-install-plan-20261002/install.sh` and `accept.sh` |

evidence/artifacts/new-wsl-install-plan-20261002/README.md

@@ -59,7 +59,7 @@ Native self-updating Claude/Codex installers remain preferred over npm; reviewed
 
 - **Rows follow the manifest.** Trafilatura, ccusage, Phoenix, Promptfoo and chezmoi lost their install and acceptance functions, mise tools, config files (`config/phoenix-compose.yaml`) and source entries; so did the rows that never had a host install (Claude Code Action, the CodeQL slot, trufflehog, the LSP plugins, the structural-diff row). attest and Dependabot became `installed: false`, with their reviewed workflow pointers kept in `notes`. Loki, Grafana and Playwright CLI keep their functions behind `--only`.
 - **Added.** ast-grep 0.45.3 through mise (`structural-search`); Alertmanager v0.34.1 as a checksum-verified release binary with `config/alertmanager.yaml` (one receiver that sends nothing), port 21093 on loopback and clustering off (`alerting`); mattpocock/skills per skill for both agents, through the `skills` installer pinned to 1.7.0, source tag v1.2.3 and telemetry off (`engineering-process-skills`). `setup-matt-pocock-skills` is included because the upstream README asks that it be one of the selected skills.
-- **Three repairs from the real run** (VALIDATION.md, findings 1 to 3): the Codex installer line sets `CODEX_NON_INTERACTIVE=1`, which the installer documents as skipping prompts; the Claude marketplace line for Trail of Bits adds the repository by its HTTPS clone URL, unpinned (the client takes a branch or tag there, not a commit, and the repository has no tag), and the acceptance prints the clone's `HEAD` beside the reviewed commit and fails when they differ; and the scripts stop with a clear message when `repo_root` is not a git checkout. The mattpocock row needs no such comparison: the installer clones the pinned tag and the acceptance checks the tag recorded in its lock file.
+- **Three repairs from the real run** (VALIDATION.md, findings 1 to 3): the Codex installer line sets `CODEX_NON_INTERACTIVE=1`, which the installer documents as skipping prompts; the Claude marketplace line for Trail of Bits adds the repository by its HTTPS clone URL, unpinned (the client takes a branch or tag there, not a commit, and the repository has no tag), and the acceptance prints the clone's `HEAD` beside the reviewed commit and fails when they differ; and the scripts stop with a clear message when `repo_root` is not a git checkout. The mattpocock row needs no such comparison: the installer clones the pinned tag. Its acceptance compares the `skillFolderHash` values that the installer recorded in its lock file with the expected git tree hashes of the six skill folders at the tag; it does not hash the installed directories. The clean run's record reports that the lock it returned had no `ref` field; the installer at v1.7.0 supports that optional field, and why it was absent in that run is unresolved.
 - **`check_plan.py`** fails and names each problem when a selected row has no install function or no post-install acceptance, a function exists for a row that is neither selected nor measurement-only, `install.sh --list` does not print exactly the rows of `install-plan.json`, a `mise.toml` tool belongs to no selected or measurement-only row, a selected row's manifest counterpart is missing, `not_installed` or `split`, two rows claim one port, or a command has no source URL. It also checks that the rows are the manifest's foundation rows, that `owners.json` agrees, that the `run_command` lines and `check` calls in the scripts are the commands in the JSON, and that every `config/` file is copied.
 
 The original [validation.json](validation.json) records the earlier source/static review and [container-validation.json](container-validation.json) the round 1 container run. The clean run of 2026-10-02 claims installation, post-install acceptance and the health of six services in a throwaway distribution, and nothing more: no provider call, model run, GPU use or sign-in, and no run on the destination distribution. After that run the header comments of both scripts and the plan's `status` text were corrected to say so; no command changed, and the record names the executed commits and file hashes.

evidence/artifacts/new-wsl-install-plan-20261002/VALIDATION.md

@@ -53,10 +53,12 @@ no provider call, no model run. The distribution was removed afterwards.
   owner, is installed by the script's first step, which prints no result line.
 - Post-install acceptance at `b48321ea`, first pass, counted in result lines, one per plan row (64): 34 exit 0, 29
   skipped and one exit 1 (`engineering-process-skills`). The 29 are the 25 rows that are not installed, the three
-  measurement-only owners and `credential-guard`, which has no host-executable check. The install was right and
-  the check was wrong: the installer's lock file records no `ref`. Each of the six skills' `skillFolderHash`
-  equals the git tree hash of its folder at tag `v1.2.3` and differs from main's, so the pin had been honoured.
-- Post-install acceptance at `1b290218`, final: the acceptance now compares those hashes. `accept.sh --only
+  measurement-only owners and `credential-guard`, which has no host-executable check. The check failed because
+  the lock that this run returned had no `ref` field; the installer at v1.7.0 supports that field as an optional
+  branch or tag, and why it was absent here is unresolved. Each of the six skills' recorded `skillFolderHash` equals
+  the git tree hash of its folder at tag `v1.2.3` and differs from main's.
+- Post-install acceptance at `1b290218`, final: the acceptance now compares those recorded hashes with the expected
+  ones; it does not hash the installed directories, and it says nothing about a client loading a skill. `accept.sh --only
   engineering-process-skills` returned exit 0, and one full `accept.sh` run returned 35 exit 0 and 29 skipped. Of
   that full run only the counts were kept, not its per-owner lines, so the record's `per_owner` table is the first
   pass and still shows this owner's failure; `per_owner_final` holds the one line that was kept.