Architecture edition: selections by repository quality (the merit rule); keys row, landed sources, distribution row

docs/decisions/2026-10-01-definitive-sota-wsl-program.md

@@ -93,6 +93,17 @@ coordinator's. One is the user's and carries a default until it is answered.
    `--configure-full-profile` unless the checkout's HEAD equals `origin/main`. A new release tag is cut after the
    merge train lands, and each receipt names the commit it installed from. Overturn: a release tag at the pins of
    record exists and the bootstrap accepts it.
+5. **What decides a selection on the new distro (the user, 2026-10-01).** The user's rule, verbatim: "make sure evl only on the reps quality itself,such as: if hindsight repo itself is better , then don't shape  the next wsl architecture into the legacy baises of our own".
+   A layer's winner on the new distro is what the repositories' own quality supports: upstream evidence and a
+   head-to-head on the same frozen tasks, each arm installed fresh by its upstream command. The source host's pins,
+   installed state and bookkeeping, and holds that came from its own integration, are not evidence for or against
+   a repository; such a hold is re-tested on the clean install. This amends decision 3 for layers that need a
+   comparison: stage 2 installs the comparison arms there, each with a fresh store, the comparison runs on the new
+   distro, and its result selects; the recorded selection is one arm and is not installed first as the default.
+   Where no head-to-head exists the merit winner is recorded as undetermined, and an incumbent is not swapped for a
+   challenger without the comparison either. The architecture edition labels its winners column as the source
+   host's selection of record. Overturn: the user restores the provisional install of the recorded selection as the
+   default.
 
 ## Phases
 

docs/decisions/2026-10-01-new-wsl-architecture-edition.md

@@ -36,12 +36,12 @@ some cells, and no run makes a layer final. The foundation rows use the first co
 us-equities rows the last, which is the basis of that assessment's synthesis. The core files those assessments cite
 (`catalogs/landscape/foundation.json`, `catalogs/saturation/ledger.json`, `manifests/stack.json`,
 `adoption/pins-linux-x86_64.json`, `catalogs/foundation/manifest.json`) are unchanged between `3361b342` and
-this edition's base. Seven pull requests carried sources that were not on main at that base. Four have merged
+this edition's base. Seven pull requests carried sources that were not on main at that base. Five have merged
 since: the new-distribution recipe (PR #569, merged as `ad7d645b`), the Harbor E2E receipt (PR #570, `000aae77`),
-the trading convergence record (PR #358, `326dc84b`) and the program record with the foundation assessment
-(PR #573, `25098f8a`). Three are open: the Codex lane's layer reviews and crosswalk (PR #575, head `51cb79ba`),
-the trading lane's verdict record (PR #578, head `e1f15c97`) and the runtime-workers blueprint README (PR #535,
-head `1195e212`).
+the trading convergence record (PR #358, `326dc84b`), the program record with the foundation assessment
+(PR #573, `25098f8a`) and the trading lane's verdict record (PR #578, `65a7b030`). Two are open: the Codex lane's
+layer reviews and crosswalk (PR #575, head `51cb79ba`) and the runtime-workers blueprint README (PR #535, head
+`1195e212`).
 
 ## Alternatives
 
@@ -84,10 +84,35 @@ trials. Upstream and user-side gates are gates.
 install default but marks it contested), `comparison_required` (from the research state or the row's owner),
 `new_host_required` (the remaining evidence needs another host) and `no_selection` (no pin of record exists).
 The research-state status maps directly; `on_requirement_change` becomes `selection_of_record_open`.
-Token efficiency is `provisional` on the Gate A owner's call. Scheduling and supervision is
+Token efficiency was `provisional` on the Gate A owner's call and reads `comparison_required` since the merit
+update below: the accepted profile is one arm and the Gate A E2E decides. Scheduling and supervision is
 `comparison_required` on the program record's call (PR #573, decision 1): its selected Dagu 2.16.6 failed the
 preregistered SIGKILL case.
 
+**The merit rule (the user, 2026-10-01).** The user's instruction, verbatim: "make sure evl only on the reps quality itself,such as: if hindsight repo itself is better , then don't shape  the next wsl architecture into the legacy baises of our own".
+The winners column of this edition was defined as the selection of record at its pin of record. That is the source
+host's bookkeeping: what the catalogs select and what this host's stack pins or installs. It is not a merit result,
+and the column is labelled as the source host's selection of record. A layer's winner on the new distribution is
+what the repositories' own quality supports: upstream evidence and a head-to-head on the same frozen tasks, each
+arm installed fresh by its upstream command. This host's pins and installed state, a missing pin, and holds that
+came from this host's own integration are not evidence for or against a repository; such a hold is re-tested on
+the clean install. Where a comparison is required the merit winner is undetermined: the row says so, names the
+arms, and the comparison's result selects. No incumbent is swapped for a challenger without that comparison, which
+would be the same bias in the other direction. What changed with this rule: the durable-memory row names
+ai-memory as the reference arm, not a default, adds agentmemory with the one matched harness result on record
+and MemPalace as arms, and no longer frames Hindsight by this host's integration holds. That result is
+configuration-level evidence, not a production head-to-head: on LongMemEval-S (Mac, 470 questions, descriptive)
+agentmemory's arm D2 scored recall_all@5 0.821 against 0.570 for ai-memory with its production embedder and the
+reranker off (arm C3); ai-memory's production arm with the LLM reranker (C4) and agentmemory's shipped-hook arm
+(D2h) have not run, the arms' captures and embedders differ, and the ai-memory build was a 2.5 pre-release.
+Also: the verdict winners that had been listed as alternatives for lacking a
+stack pin (DVC, pandera, agent-retrieval-bench, Inspect AI, MLflow) are winners of their layers, in the trading
+lane owner's words; token efficiency reads `comparison_required` with the lean base as an arm of equal standing,
+in the Gate A owner's words; and the runtime-worker candidates without a repository pin are candidates, not
+exclusions. The Codex lane's second-family reviews select different component sets in 20 of the 32 layers; under
+this rule that difference is resolved by the repositories' results in the re-record pass, with the recorded
+selection as one arm and not as the default.
+
 **Evidence classes.** The six classes of `docs/acceptance-evidence-policy.md:26-33`, plus `source_review` and
 `none_recorded`, two levels below the table: AGENTS.md keeps metadata, pinned source review and native
 execution apart. A winner's acceptance class describes a run that the cited source, or one file that source
@@ -99,10 +124,11 @@ of the 119 winner acceptance classes overstated, 13 commands that were only vers
 it could not settle; this revision corrects them. 55 of the 119 entries changed (32 classes, 48 commands and 31
 cited sources); the lane owners re-cited their winners to the files that record the runs; 13 `none_recorded`
 entries got their check to run back; restic is cited to its off-host receipt in four rows; and three rows that had
-no winner gained six. The edition now has 125 winner entries: 72 `upstream_example_or_native_operation`, 24
-`local_integration_check`, 14 `structural_validation`, 1 `independent_observation` and 14 `none_recorded`, 11 of
-which name the check to run. Row classes: 13 `local_integration_check`, 10 `none_recorded`, 7
-`structural_validation`, 6 `upstream_example_or_native_operation` and 1 `independent_observation`. The six policy
+no winner gained six; the merit update added five verdict winners. The edition now has 130 winner entries: 73
+`upstream_example_or_native_operation`, 24 `local_integration_check`, 14 `structural_validation`, 1
+`independent_observation` and 18 `none_recorded`, 15 of which name the check to run. Row classes: 13
+`none_recorded`, 12 `local_integration_check`, 6 `upstream_example_or_native_operation`, 5
+`structural_validation` and 1 `independent_observation`. The six policy
 classes have no defined order, so the build enforces what can be enforced:
 when any winner's acceptance is `none_recorded`, the row's class is `none_recorded`; otherwise the row's class is
 one that at least one winner's acceptance carries, and where the winners carry several, the edition records the
@@ -131,12 +157,14 @@ reordered research state fails the build instead of showing each row's states be
 without a row is listed on the page, not a build failure, so adding a layer elsewhere never breaks the page.
 
 **This edition's rows.** 37 rows: 20 foundation, 12 us-equities, 5 cross. Closed: none. 20
-`selection_of_record_open`, 14 `comparison_required`, 2 `new_host_required`, 1 `provisional` and no
+`selection_of_record_open`, 15 `comparison_required`, 2 `new_host_required`, and no `provisional` or
 `no_selection`. Every assessed row has at least one closure item unmet or partial, and no review item
 (c4) is met. The twelve us-equities rows follow the trading closure assessment: their winners are the components
 the layer's current choice names that have a pin of record (runtime-target.json for the engine, the brokers
-and the adaptive paper engine; stack pins for profile components), while the 2026-09-22 verdict winners
-without such a pin (DVC, pandera, agent-retrieval-bench, Inspect AI, MLflow) stay alternatives.
+and the adaptive paper engine; stack pins for profile components). The 2026-09-22 verdict winners without
+such a pin (DVC, pandera, agent-retrieval-bench, Inspect AI, MLflow) were first listed as alternatives; under the
+merit rule they are winners of their layers at the verdict's baseline version, because a missing stack pin means
+not installed on the source host, never not selected.
 The trading lane owner's verdicts of 2026-10-01 call every trading layer "selection of record, open". Two
 layers had no selection of record for the layer itself (research-factors-ml and security-supply-chain). After a
 cross-family adjudication in the Codex lane's reviews, the owner recorded one for each the same day (PR #578):
@@ -145,16 +173,19 @@ second, which reads `selection_of_record_open`, with Grype pinned by the foundat
 decision closes anything. The other rows read `comparison_required` where the research state asks for a
 comparison and `selection_of_record_open` otherwise. The owner corrected the trading winners' acceptance sources
 in this change's pull request (#574) and re-read the twelve rows there, and each row's notes carry the owner's
-verdict and blocking item verbatim. The keys lane corrected the credential-practice row, and the Gate A owner
-decided the acceptance entries of its three rows the same way. The backtesting-engine row carries the dispute
-recorded at PR #358's head `b0eb7a11`.
+verdict and blocking item verbatim. The keys lane corrected the credential-practice row and worded it again
+after its canary proof merged (PR #579, the first update of this edition under overturn condition 5), and the
+Gate A owner decided the acceptance entries of its three rows the same way. The backtesting-engine row carries
+the dispute recorded at PR #358's head `b0eb7a11`.
 Paper results appear only as fills and passed trials (the 2026-09-29 Alpaca series: 11 of 13 trials passed,
 41 entries and 39 exits filled). The cross row for the distribution is `new_host_required`: its image, hash
 and creation path are on main since PR #569, selected by source review, and nothing has run on a host; the
-recipe's follow-up with the completeness critic's pre-checks is open, and stage 1 waits for it.
+recipe's follow-up with the completeness critic's pre-checks merged with PR #582 (`b8dd81dd`), so the page now
+starts with a rehearsal on a throwaway name and pre-checks in the workstation distribution.
 
-**Install order on the new distribution** (row `cross:wsl-distro`, each step citing the recipe of PR #569):
-stage 1 on Windows; first boot (systemd, linger, user bus, packages, subordinate ids, the login hand-off);
+**Install order on the new distribution** (row `cross:wsl-distro`, each step citing the recipe of PR #569
+and #582): the rehearsal and the pre-checks; stage 1 on Windows; first boot (systemd, linger, user bus, packages,
+subordinate ids, the login hand-off);
 a clone of origin/main and the host file on free ports; credentials before any keyed tool; stage 2 with the
 token-efficiency profile, native sign-in, the Context Mode plugin started once before the settings step
 (its start script writes the SessionStart cache-heal hook that the settings template runs), then
@@ -205,8 +236,8 @@ Write a new edition, or update this one in the change that causes it, when any o
    in the Linux pins file or runtime-target.json is caught by the line-bounds check only when the line
    disappears, so re-read those rows.
 3. **A new upstream release** of a winner, or a qualification that moves a pin (Codex 0.159.3 first).
-4. **A pending source lands.** PR #575, #578 or #535 merges: the build hashes the landed file and labels it as
-   landed after this edition's base. PR #569, #570, #358 and #573 have landed: their citations are plain
+4. **A pending source lands.** PR #575 or #535 merges: the build hashes the landed file and labels it as
+   landed after this edition's base. PR #569, #570, #358, #573 and #578 have landed: their citations are plain
    `source_path` entries now, and `cross:wsl-distro` was re-rated when #569 did.
 5. **A lane owner changes a verdict or its wording** for one of its rows.
 6. **The Gate A multi-agent E2E runs** on the new distribution: it decides the token-efficiency verdict.
@@ -235,8 +266,8 @@ Write a new edition, or update this one in the change that causes it, when any o
 - Docker Engine documentation, rootless mode, prerequisites and troubleshooting (read 2026-10-01); the Next.js
   v16.3.8 release page (read 2026-10-01)
 - PR #358 (merged as `326dc84b`): `catalogs/us-equities/convergence-20260926.json`
-- PR #578 at `e1f15c97`: `docs/decisions/2026-10-01-trading-layer-verdicts.md` (the trading lane owner's verdicts
-  and the two selections of 2026-10-01)
+- PR #578 (merged as `65a7b030`): `docs/decisions/2026-10-01-trading-layer-verdicts.md` (the trading lane
+  owner's verdicts and the two selections of 2026-10-01)
 - Upstream release pages read on 2026-10-01 (07:41Z to 08:06Z), each cited on its winner; the
   nautilus_trader issue #4983 and pull request #5041 states; the gitleaks README; the Harbor README
   (`uv tool install harbor`)

docs/ecosystem/README.md

@@ -137,7 +137,8 @@ date, its base commit, scope, the verdict rules verbatim, what each verdict and
 evidence class means, the five closure items of
 [the research state](../../catalogs/landscape/research-state.json) and the
 edition's sources, and states how many rows are closed. One table per catalog
-lists each row's winners at their pin of record, verdict, evidence class,
+lists each row's source-host selection of record (the winners column, which is
+this host's bookkeeping and not a merit result) at its pin, verdict, evidence class,
 reasons and install command; expanding a row shows the closure items with what
 is missing, the winners' pin locators, install and acceptance commands and
 upstream currency, the alternatives, the ordered new-host steps and the gates.

docs/ecosystem/template.html

@@ -121,7 +121,7 @@
 <p class="subnote">Codex input includes its cached-input subset. Claude ordinary input, cache creation and cache reads use different categories. Do not add those conventions blindly or double-count the same telemetry receipt.</p>
 </section>
 <section id="architecture" role="tabpanel" aria-labelledby="tab-architecture" hidden>
-<div class="page-heading"><div><div class="eyebrow">New WSL distro / one dated edition</div><h2>The final architecture, layer by layer.</h2><p>Each row names the selection of record at its recorded pin, the verdict under the edition's rules, the evidence class, the reasons with their sources and what a session runs on the new distro. Expand a row for the five closure items, alternatives, upstream currency, ordered new-host steps and gates. A new distro collects its own evidence; nothing here is acceptance on it.</p></div></div>
+<div class="page-heading"><div><div class="eyebrow">New WSL distro / one dated edition</div><h2>The final architecture, layer by layer.</h2><p>Each row names the source host's selection of record at its recorded pin (not a merit result: a comparison row's winner is undetermined until its arms run on the new distro), the verdict under the edition's rules, the evidence class, the reasons with their sources and what a session runs on the new distro. Expand a row for the five closure items, alternatives, upstream currency, ordered new-host steps and gates. A new distro collects its own evidence; nothing here is acceptance on it.</p></div></div>
 <div id="architecture-edition" class="scope-note landscape-summary"></div>
 <div id="architecture-actions" class="landscape-actions"></div>
 <div id="architecture-topic"></div>
@@ -797,7 +797,7 @@
   const closure = make("dl", "catalog-facts");
   edition.closure_items.forEach(item => add(closure, make("dt", "", item.id + " · " + readable(row.closure[item.id])), make("dd", "", item.text)));
   add(detail, make("h4", "", "Closure items"), closure, make("p", "", row.closure.missing ? "Missing: " + row.closure.missing : "Nothing missing: all five closure items hold."));
-  if(row.winners.length) add(detail, make("h4", "", "Winners: pin of record, install, acceptance and upstream currency"));
+  if(row.winners.length) add(detail, make("h4", "", "Source host's selection of record: pin, install, acceptance and upstream currency"));
   row.winners.forEach(winner => {
     const block = add(make("div", "catalog-detail"), make("strong", "", [winner.label, winner.role, winner.pin].filter(Boolean).join(" · ")), add(make("p", "small"), link("Pin of record: " + winner.pin_source.locator + " ↗", winner.pin_source.url)));
     if(winner.pin_drift !== null) add(block, make("p", "small", architectureDrift(winner)));
@@ -853,7 +853,7 @@
     const subset = rows.filter(row => row.catalog === catalog);
     if(!subset.length) return;
     const table = make("table", "token-table"), body = make("tbody");
-    add(table, add(make("thead"), add(make("tr"), ...["Layer", "Winners (pin)", "Verdict", "Evidence class", "Reasons", "Install on the new distro"].map(label => make("th", "", label)))));
+    add(table, add(make("thead"), add(make("tr"), ...["Layer", "Source host's selection (pin)", "Verdict", "Evidence class", "Reasons", "Install on the new distro"].map(label => make("th", "", label)))));
     subset.forEach(row => {
       const layer = add(make("td"), make("strong", "", row.title), make("p", "small mono", row.catalog + "/" + row.layer_id));
       if(row.research_status) add(layer, make("p", "small muted", "Research state: " + readable(row.research_status)));

manifests/evidence.json

@@ -15610,8 +15610,8 @@
     },
     {
       "path": "docs/ecosystem/README.md",
-      "sha256": "d13d2e71db5bbf79f9133a34ffd109ba0217cb59e6fe6099f85e815998b062c1",
-      "bytes": 17975
+      "sha256": "ed4b5384848386d0a6077cd17147c2dd02dc7aeaf463ec8807c3b3ba0effb21b",
+      "bytes": 18065
     },
     {
       "path": "docs/ecosystem/claude-repository-evidence.html",
@@ -15650,8 +15650,8 @@
     },
     {
       "path": "docs/ecosystem/template.html",
-      "sha256": "c61e23df300b0872bc784d5dc7926ef23009ba6d9f16acc4ae9e0f0aab8ae7d3",
-      "bytes": 125994
+      "sha256": "b1fc23b3595dca2164f0b3fc7bb19d553caa0be8c4fe4b25e70004a48ebcc165",
+      "bytes": 126141
     },
     {
       "path": "docs/evidence.md",
@@ -46920,8 +46920,8 @@
     },
     {
       "path": "tests/test_ecosystem_manifest.py",
-      "sha256": "8e0fbb9b579e99eb740155a9e728c43146ac61ce51fa0bbce47a344482822561",
-      "bytes": 137469
+      "sha256": "d1990f63256a86d69eaa154217aba92548905ff0bafc61e574465425e9ba62bb",
+      "bytes": 137485
     },
     {
       "path": "tests/test_effort_default_guard.py",