docs(observatory): record selector rules and local UI acceptance

docs/OBSERVATORY_REALTIME_PROFILES_EXECPLAN.md

@@ -16,6 +16,9 @@ Docker-профилей, затем переносим выбранный про
 версий остаются в «Лабораторных доказательствах». Если рассчитаны все текущие
 совместимые профили, выбор пуст/недоступен, кнопки «Рассчитать» нет, «Обновить»
 остаётся. Старое имя LAB не доказывает совпадение версии конфигурации.
+Уточнение владельца 2026-09-03: никаких кнопок/плашек «Расчёт завершён»,
+«Результат опубликован» и аналогичных подтверждений. Готовность видна по
+результату с маркером профиля внизу; в выборе остаются только нерассчитанные.
 
 ### Этап 1 — Идентичность расчёта и переиспользование результата (в работе)
 
@@ -86,7 +89,7 @@ succeeded/not-required не принят за новый результат. В
 сохранены. Worker/model jobs не создавались. Browser: каталог, выбор, Refresh,
 normal/expanded и Escape проверены, console warnings/errors0.
 
-Этап1 пользовательски ещё не закрыт: текущий селектор смешивает архивные entries
+На момент активации cache этап1 пользовательски ещё не был закрыт: селектор смешивал архивные entries
 и portable definitions, а source/setup-only выбор последней job показывает
 старое «Расчёт завершён» рядом с новым «Рассчитать». Это не текущий cache hit.
 Следующий шаг на стыке этапов1–2: exact-definition queue/status binding,
@@ -95,6 +98,27 @@ normal/expanded и Escape проверены, console warnings/errors0.
 Подробности/rollback в том же отчёте выше; повторять сборку без изменения кода
 или снова освобождать память не требуется.
 
+**Четвёртый инкремент, 2026-09-03: правило выбора исправлено.** Селектор получает
+только portable definitions, оставляет совместимые без проверенного exact cache
+hit; архивные LAB и рассчитанные версии остаются в доказательствах ниже.
+Кнопки/плашки завершения и дублирующее «Записанный разбор» убраны. Все загруженные
+результаты показаны без прежнего лимита6. «Рассчитать» отсутствует при пустом
+выборе и во время уже принятого расчёта; до успешной проверки запуска кнопка
+неактивна. Во время работы есть только индикатор ожидания, без вымышленных
+процентов. Ошибки и retry публикации сохранены вне строки выбора.
+Queue GET фильтрует точную definition SHA до LIMIT; при смене source/setup/SHA
+предыдущие jobs/errors не показываются даже до React effects. Публикация
+обновляет оба каталога; удаление projection также обновляет выбор.
+720 frontend tests,52 focused queue/API tests, architecture4/4, typecheck,
+Ruff/mypy и production build PASS. Подробная финальная runtime/UI-проверка —
+в [отчёте](../experiments/perception/OBSERVATORY_PUBLISHED_CACHE_2026-09-03.md).
+Визуальная проверка также выявила сжатие селекторов в узком обычном окне:
+существующая адаптивная раскладка теперь зависит от ширины окна Observatory,
+а не всего viewport. Старые terminal logs не выдаются за ошибку нового выбора.
+Полный новый LAB-run, реальные счётчики прогресса и positive cache/view на
+операторских данных остаются следующим acceptance этапа2, а не объявляются
+закрытыми этой правкой интерфейса.
+
 - Сверить каталог, очередь, публикацию и просмотр для текущих M4.9T5 и LAB V1.
 - Связать готовность с точной исходной записью и immutable RunDefinition,
   включающей image/model/config/adapter/result-contract identities.

docs/adr/0050-recorded-observatory-first.md

@@ -121,6 +121,28 @@ binding and separating archives from executable choices remain required;
 the activated backend cache does not prove those UI changes complete.
 See [verification and activation status](../../experiments/perception/OBSERVATORY_PUBLISHED_CACHE_2026-09-03.md).
 
+## Selector contract clarification — 2026-09-03
+
+The owner explicitly excludes completion buttons and status badges. A completed
+current profile is represented by its reviewable, profile-marked evidence below,
+not by a “calculation complete” control. The selector now contains compatible
+portable definitions with no verified exact published cache hit. An archived LAB
+or old definition never hides a new version. An empty selector has no Calculate.
+An in-flight computation has only an activity indicator; failed publication
+keeps a separate recovery action and never starts inference again.
+
+Removed the six-card presentation cap; the existing catalog window limit remains
+explicit. Queue reads now filter definition SHA before pagination. Hook snapshots
+are fenced by source/setup/definition so previous jobs cannot flash in the next
+selection. New synthetic tests exercise none/all/partially cached choices,
+unavailable executors, legacy/new-version separation and pre-effect selection
+changes. Sequential validation:720 frontend tests,52 queue/API tests, architecture,
+typecheck, Ruff/mypy and production build PASS. Full recorded-run/product
+acceptance and measured progress counters remain stage2.
+Historical terminal logs do not become a new selection's error. The existing
+responsive selector layout follows workspace width, including restored narrow
+windows, rather than only the browser viewport.
+
 ## Boundaries
 
 No Synology deployment, hardware actuation, motor integration, new capture,

experiments/perception/OBSERVATORY_PUBLISHED_CACHE_2026-09-03.md

@@ -137,7 +137,7 @@ Sequential validation under `mission-core-product-ui`:
 Private snapshots/logs and previous frontend: `.runtime/docker-vm-recovery-z1636o/`.
 No raw recording content was collected into this diagnostic evidence.
 
-## Remaining product acceptance
+## Product acceptance remaining at initial cache activation
 
 The cache increment is active, but the complete workflow is not accepted.
 Browser QA reproduced the next-stage UI mismatch: archived entries and portable
@@ -151,3 +151,70 @@ progress, complete recorded calculation and sequential M4.9T5/LAB V1 acceptance.
 Stage1 user-facing reconciliation and the full stage2 workflow remain open.
 Do not repeat the memory recovery or claim/v3 repair, or treat the old realtime
 FAIL as a new gate. No per-LAB UI branch or automatic inference on Refresh.
+
+## Selector correction and acceptance — 2026-09-03
+
+The owner clarified that a finished calculation is represented only by its
+profile-marked evidence below, never by a completion button/status. Implemented
+in the shared Observatory surface using existing Design Guideline components:
+
+- Only compatible, uncalculated portable definitions are offered. The verified
+  catalog cache decision remains authoritative; names, archived definitions,
+  failed or unpublished jobs cannot hide a current profile. No Calculate when
+  the selection is empty. During an accepted computation only an activity
+  indicator appears; no invented percentage. Before admission the button is
+  disabled. Failed publication retains recovery without new inference.
+- Removed completion badges, the duplicate recorded-review badge and the
+  six-result presentation cap. All results in the loaded catalog are shown.
+  The existing catalog-window bound remains explicit; this is not complete
+  pagination for500 recordings.
+- Queue reads filter source/setup/definition before LIMIT. UI snapshots are
+  fenced by the same selection key before React effects, so previous jobs and
+  request errors cannot flash against a new selection. Historical terminal logs
+  are not displayed as new operator failures; observed/current failures remain
+  actionable.
+- Publication refreshes both catalogs; projection deletion refreshes selection.
+  Browser QA found selector compression in a restored narrow window; moved the
+  existing responsive layout from viewport width to workspace container width.
+  No new component, per-LAB page, model configuration or Worker deployment.
+
+Sequential checks: architecture4/4, full typecheck, frontend720/720,
+focused queue/API52/52, Ruff and mypy PASS. A synthetic test proves definition
+filtering happens before pagination; API accepts a valid digest and rejects a
+malformed one with422. Tests cover zero/all/partly cached choices, old versions,
+incompatible/archived entries, missing projections and stale first-render data.
+Final production build14.56s, peak RSS2108293120bytes; unchanged large-chunk
+warnings. Pressure remained1 and swap fell2666.19→2618.19MiB. No concurrent heavy
+tests/builds/browser runs; Docker limits/containers and operator apps untouched.
+
+Active canonical8000: backend PID69976, health200/reconciler ready, frontend
+`/assets/index-BHcT8phv.js`, CSS `/assets/index-pPpyKjkx.css`,
+index SHA256 `97da43251ef0d0635eb078ee1a102fa820f1716bb9ec5879dffb284aa3628ea6`.
+No alternate backend8765/preview4173. The existing LaunchAgent restarted only
+the canonical backend for the API change; final frontend-only corrections did
+not restart it again. Original frontend backup:
+`.runtime/observatory-selector-fQUwPT/frontend-dist-previous`.
+For rollback, restore that build and the matching pre-change code on the same
+8000 endpoint. Do not replace or truncate the queue DB.
+
+In-app browser acceptance:
+
+- 004TREE offers exactly the two current portable profiles, not four mixed
+  archive/current entries. Both select correctly; Calculate becomes available
+  after preflight. Completion labels and historical error alerts are absent.
+- 01 has an empty disabled selector and no Calculate because its capture
+  attestation is missing. This is incompatibility, not a fake cache hit.
+- 00 renders all14 loaded evidence entries. 004TREE retains both old entries
+  and its existing admitted LAB V1 review action; no archive projection was
+  deleted or falsely promoted into a current verified result.
+- Refresh, dropdown Escape, normal/expanded windows and readable selector text
+  verified. Final browser console warnings/errors0; completion labels0,
+  alerts0. No Calculate, publication retry, delete, rename or replay was invoked.
+
+All six queue-table row hashes are identical before/after this correction and
+browser checks:11 jobs,50000 legacy receipts,1 reconciliation,0 live leases,
+v3 grants and preemptions. No inference or publication jobs were created.
+Current real data still contains no exact successful/published cache hit;
+all-cached/partial-cache behavior is synthetic-test evidence, not a new completed
+LAB run. The remaining stage2 work is measured progress, full sequential
+M4.9T5/LAB V1 calculation, publication, and cached viewing after restart.