fix(merge-gate): tolerate re-test infra-timeout + tree-kill spawned pytest

Eliminate the false `deploy-staging -> development` rollback that fired when the
merge-gate local re-test timed out (infra/resource) on a green CI + tester +
staging branch (incident ORCH-109/PR #129: a 516.7s suite blew its 600s budget
under CPU starvation from orphaned pytest processes -> timeout misrouted as a
code fault -> developer-retry loop -> manual gate).

Additive, 5 independent kill-switches, never-raise, self-hosting scope. Untouched
byte-for-byte: STAGE_TRANSITIONS, the QG_CHECKS registry, check_branch_mergeable
name/semantics, machine-verdict keys, the DB schema. INV-4 (never push/force-push
main) and the no-prod-restart rule are preserved.

- D1: new stdlib-only leaf src/proc_group.py runs the spawned re-test/coverage
  pytest in its own process group (start_new_session) and tree-kills the WHOLE
  group on timeout (os.killpg SIGTERM->grace->SIGKILL); used by
  merge_gate.retest_branch and coverage_gate.measure_coverage. No orphan leak.
  Fallback never-break: subprocess_tree_kill_enabled=False / non-POSIX -> the
  prior subprocess.run.
- D2/D3: merge_gate.classify_retest_failure distinguishes timeout/red/lock-busy/
  other; an infra timeout routes to _handle_merge_gate_infra_retry (bounded
  re-queue, task stays on deploy-staging, no rollback / no developer-retry); a
  red re-test / conflict still rolls back (BR-6). Exhaustion -> one infra alert.
- D4: skip the local re-test when the pre-merge rebase was a proven no-op (HEAD
  already CI/tester/staging-validated); fail-safe runs the re-test on any
  uncertainty. Flag merge_retest_skip_when_current_enabled.
- D5: merge_retest_timeout_s 600 -> 900 + _resolve_retest_timeout validation;
  reaper_max_running_s invariant preserved without change.
- D6: in-process counters + read-only merge_gate block in GET /queue; appended
  ("ORCH-110","classify_retest_failure","src/merge_gate.py") to
  MAIN_REGRESSION_MARKERS. Docs (README/internals overview/CLAUDE/CHANGELOG/
  .env.example) updated in the same PR.

Tests: tests/test_orch110_*.py (TC-01..TC-12, incl. the red-before/green-after
incident regression). Full suite green (1988 passed).

Refs: ORCH-110

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.env.example

@@ -164,11 +164,32 @@ ORCH_TRACKER_LIVE_STATUS_TIMEOUT_S=3
 #   DEFER_MAX_ATTEMPTS -> defer retries before escalation (avoids livelock).
 ORCH_MERGE_GATE_ENABLED=true
 ORCH_MERGE_GATE_REPOS=
-ORCH_MERGE_RETEST_TIMEOUT_S=600
+# ORCH-110 (D5): re-test budget raised 600 -> 900 (74% headroom over the observed
+# 516.7s suite). Cross-invariant (ORCH-065/109): keep ORCH_REAPER_MAX_RUNNING_S
+# (5400) > Σ(deploy-staging gate-work) + grace if you raise this — see
+# docs/work-items/ORCH-110/07-infra-requirements.md.
+ORCH_MERGE_RETEST_TIMEOUT_S=900
 ORCH_MERGE_RETEST_TARGET=tests/
 ORCH_MERGE_LOCK_TIMEOUT_S=300
 ORCH_MERGE_DEFER_DELAY_S=60
 ORCH_MERGE_DEFER_MAX_ATTEMPTS=5
+# ORCH-110: merge-gate re-test infra-timeout tolerance + tree-kill of the
+# orchestrator-spawned pytest subprocess (re-test + coverage). Each default = the
+# desired prod behaviour; each flag is an independent kill-switch (off ->
+# byte-for-byte pre-ORCH-110). The tree-kill grace reuses ORCH_AGENT_KILL_GRACE_SECONDS.
+#   SUBPROCESS_TREE_KILL_ENABLED          -> D1: spawn re-test/coverage pytest in its
+#       own process group; tree-kill the WHOLE group on timeout (no orphan grandchildren).
+#   MERGE_RETEST_INFRA_TOLERANCE_ENABLED  -> D3: a re-test TIMEOUT is a transient
+#       (bounded infra-retry, NOT a code-fault rollback to development).
+#   MERGE_RETEST_INFRA_MAX_RETRIES        -> D3: infra-retry budget before an infra-alert.
+#   MERGE_RETEST_INFRA_RETRY_DELAY_S      -> D3: delay before the staging-deployer re-run.
+#   MERGE_RETEST_SKIP_WHEN_CURRENT_ENABLED-> D4: skip the local re-test when the
+#       pre-merge rebase was a proven no-op (HEAD already CI/tester/staging-validated).
+ORCH_SUBPROCESS_TREE_KILL_ENABLED=true
+ORCH_MERGE_RETEST_INFRA_TOLERANCE_ENABLED=true
+ORCH_MERGE_RETEST_INFRA_MAX_RETRIES=2
+ORCH_MERGE_RETEST_INFRA_RETRY_DELAY_S=120
+ORCH_MERGE_RETEST_SKIP_WHEN_CURRENT_ENABLED=true
 # ORCH-026 Level A: unconditional pre-merge rebase. With the flag ON (default),
 # check_branch_mergeable ALWAYS rebases the branch onto origin/main under the held
 # merge-lease (not only when behind) — a deterministic structural anti-phantom on

.task-dev.md

@@ -1,4 +1,4 @@
-Work item: ORCH-111
+Work item: ORCH-110
 Repo: orchestrator
-Branch: feature/ORCH-111-bug-watchdog-must-alert-on-lon
+Branch: feature/ORCH-110-bug-merge-gate-local-re-test-t
 Stage: development

CHANGELOG.md

@@ -3,6 +3,13 @@
 Формат: [Keep a Changelog](https://keepachangelog.com/). Записи — на смысловой PR/задачу.
 
 ## [Unreleased]
+- **Merge-gate re-test: толерантность к инфра-таймауту + tree-kill спавненных pytest + контракт необходимости re-test** (ORCH-110, `fix`, bug→escalate full-cycle): устранён ложный откат `deploy-staging → development`, возникавший когда локальный re-test merge-gate падал по **таймауту** (инфра/ресурс) при зелёных CI + tester + staging (инцидент ORCH-109/PR #129: сюит 516.7s упёрся в бюджет 600s под CPU-голоданием от осиротевших pytest-процессов → `(False, "re-test timeout after 600s")` → `_handle_merge_gate_rollback` → каждый из 3 developer-retry падал так же → «Merge-gate still failing after 3 developer retries» → ручное вмешательство). Аддитивно, под 5 независимыми kill-switch, never-raise, скоуп self-hosting; `STAGE_TRANSITIONS`/реестр `QG_CHECKS`/семантика `check_*`/machine-verdict-ключи/схема БД — **байт-в-байт не тронуты**; INV-4 (никогда push/force-push `main`) и запрет рестарта прод-контейнера — соблюдены. ADR: `docs/work-items/ORCH-110/06-adr/ADR-001-merge-gate-retest-infra-tolerance-and-tree-kill.md`, сквозной `docs/architecture/adr/adr-0042-merge-gate-retest-infra-tolerance-and-tree-kill.md`.
+  - **D1 — process-group tree-kill (`src/proc_group.py`, новый stdlib-only leaf):** `merge_gate.retest_branch` и `coverage_gate.measure_coverage` теперь спавнят pytest в **отдельной группе процессов** (`start_new_session`) и при таймауте убивают **всё дерево** (`os.killpg`, каскад SIGTERM→grace→SIGKILL по образцу `launcher.stop_process`), а не только прямого потомка — осиротевшие внуки-pytest больше не переживают бюджет и не грузят CPU. Контракты возврата сохранены (меняется лишь побочный эффект — нет утечки). Грейс реюзит `agent_kill_grace_seconds`. Fallback never-break: `subprocess_tree_kill_enabled=False` или не-POSIX → прежний `subprocess.run(timeout=)`.
+  - **D2/D3 — классификация + маршрутизация инфра-таймаута:** чистый предикат `merge_gate.classify_retest_failure(reason)` различает `timeout`/`red`/`lock-busy`/`other` (scope-guard: `auto_rebase_onto_main`'s «rebase timeout» — НЕ инфра-таймаут re-test, остаётся на rollback-пути). Инфра-таймаут → новый `_handle_merge_gate_infra_retry` (ограниченный повтор по образцу `_handle_merge_gate_defer`: задача **остаётся на deploy-staging**, staging-deployer перезапускается с задержкой, **БЕЗ** отката на `development` и **БЕЗ** расхода developer-retry). Анти-над-толерантность (BR-6): детерминированно **красный** re-test / конфликт по-прежнему → `_handle_merge_gate_rollback`. Anti-loop: исчерпание бюджета → один **инфра-alert** (явно инфраструктурная формулировка «НЕ дефект кода» с кликабельным номером), задача НЕ уходит в `development`.
+  - **D4 — контракт необходимости re-test:** при `premerge_rebase_always=True` re-test теперь **пропускается**, когда rebase оказался доказанным no-op (HEAD не сдвинулся = ветка уже содержит свежий `origin/main`, тот же коммит уже подтвердили CI + tester + staging) — distribute той же оптимизации, что путь `premerge_rebase_always=False` уже имеет для не-behind ветки. Fail-safe: при любой неопределённости (`head_sha` пуст / git-ошибка) re-test **выполняется** (BR-6/AC-3 не ослаблен).
+  - **D5 — бюджет:** `merge_retest_timeout_s` 600 → **900** (запас 74% над 516.7s) + валидация `_resolve_retest_timeout` (малформ/непозитив → дефолт 900 + WARNING). Сквозной инвариант ORCH-065/109 `reaper_max_running_s (5400) > Σ(deploy-staging gate-work)+grace (≈4460)` соблюдён **без** правки `reaper_max_running_s`.
+  - **D6 — наблюдаемость:** in-process счётчики (`retest_timeout_total`/`retest_infra_retry_total`/`retest_infra_exhausted_total`/`retest_skipped_current_total`/`last_infra_timeout_wi`) + read-only блок `merge_gate` в `GET /queue` (отличим от код-фейл-отката); координация с ORCH-111 (`proc_blocking`) без дубля (ORCH-110 предотвращает/толерирует, ORCH-111 наблюдает). Append-only regression-guard: добавлен `("ORCH-110", "classify_retest_failure", "src/merge_gate.py")` в `MAIN_REGRESSION_MARKERS`.
+  - **Конфиг (5 новых ключей, дефолт = боевое):** `ORCH_SUBPROCESS_TREE_KILL_ENABLED`/`ORCH_MERGE_RETEST_INFRA_TOLERANCE_ENABLED`/`ORCH_MERGE_RETEST_INFRA_MAX_RETRIES=2`/`ORCH_MERGE_RETEST_INFRA_RETRY_DELAY_S=120`/`ORCH_MERGE_RETEST_SKIP_WHEN_CURRENT_ENABLED` + бамп `ORCH_MERGE_RETEST_TIMEOUT_S=900`. Покрытие — `tests/test_orch110_*.py` (TC-01…TC-12, включая регресс инцидента red-before/green-after).
 - **Watchdog-сигнал `proc_blocking`: алерт на долго живущий осиротевший тест-процесс** (ORCH-111, `feat`): закрыта слепая зона наблюдаемости между `agent_hung` (видит только треканые джобы по `jobs.pid`) и осиротевшими субпроцессами `pytest`, которые орк запускает сам (`merge_gate.retest_branch`/`coverage_gate.measure_coverage`) и которые при timeout-kill агента (`-9`, ORCH-109) репарентируются на tini и живут сутками, грузя CPU и валя merge-gate re-test (инцидент: процессы `test_install_lite_script.py` жили >2 суток без единого алерта). Изменения **строго внутри наблюдателя** (`watchdog/**` + сервис watchdog в compose); `src/**`/`/metrics`/`schema_version`/`STAGE_TRANSITIONS`/`QG_CHECKS`/`check_*`/machine-verdict/схема БД — **байт-в-байт не тронуты**; выкат пересобирает **только** `orchestrator-watchdog`, прод `orchestrator` не рестартится (NFR-3). ADR: `docs/work-items/ORCH-111/06-adr/ADR-001-watchdog-orphan-test-process-alert.md`, сквозной `docs/architecture/adr/adr-0041-watchdog-orphan-test-process-alert.md`.
   - **Коллектор `watchdog/collectors/proc.py` (D3):** новый stdlib-only `/proc`-скан (под `pid: host` контейнерный `/proc` отражает хост-namespace) — читает `/proc/stat` (`btime`) + `os.sysconf("SC_CLK_TCK")`, итерирует числовые `/proc/<pid>`, матчит `/proc/<pid>/cmdline` по паттерну тест-класса, парсит `/proc/<pid>/stat` (поле 22 `starttime` → `age_s`, поля 14+15 `utime+stime` → `cpu_s` информационно). Строго **read-only** (никаких `os.kill`/сигналов/`subprocess`; **никогда** не читает `/proc/<pid>/environ` — секреты); **never-raise** (per-pid гонка «процесс умер между listdir и read» пропускается, top-level → `[]`); чистый разбор отделён от I/O (тестируется на фейковом `/proc`-дереве).
   - **Чистый builder `proc_signals` + синтез RECOVERY (D4):** per-entity `Signal("proc_blocking", pid)` active ⇔ `age_s > cfg.proc_age_s` (cmdline уже отфильтрована коллектором); действенный RU-`detail` (PID + возраст + усечённый фрагмент cmdline + CPU-время). Исчезновение процесса не оставляет «висящего» алерта: в `core.tick()` для каждого alerting-ключа без свежего сигнала **синтезируется** `Signal(active=False)` → существующая `decision.decide()`/`AlertState` даёт **однократный** RECOVERY и чистит состояние (никакой новой анти-спам-логики — FR-5).

CLAUDE.md

@@ -235,6 +235,54 @@ kill-switch, never-raise, fail-safe → полный цикл.
   `docs/work-items/ORCH-027/06-adr/ADR-001-coverage-gate.md`,
   `docs/architecture/adr/adr-0029-coverage-gate.md`.
 
+## Merge-gate re-test: инфра-толерантность + tree-kill + контракт re-test (ORCH-110)
+Багфикс инцидента **ORCH-109/PR #129** (bug → escalate full-cycle): локальный re-test merge-gate
+(`check_branch_mergeable`, ребро `deploy-staging → deploy`) падал по **таймауту** (516.7s-сюит упёрся
+в бюджет 600s под CPU-голоданием от **осиротевших** pytest-процессов) при зелёных CI+tester+staging →
+маршрутизировался как **код-фейл** в `_handle_merge_gate_rollback` (откат на `development` + расход
+developer-retry) → каждый retry падал так же → «Merge-gate still failing after 3 developer retries» →
+ручное вмешательство. Аддитивно, под **5 независимыми kill-switch**, never-raise, скоуп self-hosting;
+`STAGE_TRANSITIONS`/реестр `QG_CHECKS`/семантика/имя `check_branch_mergeable`/machine-verdict-ключи/
+схема БД — **байт-в-байт не тронуты** (merge-gate остаётся под-гейтом-врезкой, не новой стадией/QG);
+INV-4 (никогда push/force-push `main`) и запрет рестарта прод-контейнера — соблюдены.
+- **D1 (tree-kill, корень утечки):** новый stdlib-only leaf `src/proc_group.py`
+  (`run_in_process_group`, паттерн чистоты `serial_gate`) спавнит спавненный pytest в **отдельной
+  группе процессов** (`start_new_session`) и при таймауте убивает **всё дерево** (`os.killpg`, каскад
+  SIGTERM→grace→SIGKILL, зеркало `launcher.stop_process`, грейс = `agent_kill_grace_seconds`).
+  Используют его `merge_gate.retest_branch` И `coverage_gate.measure_coverage` (сиблинг-источник
+  утечки). Контракты возврата сохранены — меняется лишь отсутствие сирот. Fallback never-break:
+  `subprocess_tree_kill_enabled=False`/не-POSIX → прежний `subprocess.run`.
+- **D2/D3 (классификация + маршрутизация):** чистый предикат `merge_gate.classify_retest_failure`
+  различает `timeout`/`red`/`lock-busy`/`other` (scope-guard: «rebase timeout» git'а — НЕ инфра-таймаут
+  re-test). Инфра-таймаут → `_handle_merge_gate_infra_retry` (зеркало `_handle_merge_gate_defer`:
+  задача **остаётся на `deploy-staging`**, staging-deployer перезапускается с задержкой, **БЕЗ** отката
+  на `development` и **БЕЗ** developer-retry; restart-safe счётчик `_merge_infra_retry_count` по маркеру
+  `merge-gate infra-timeout retry` в `task_content`). Анти-над-толерантность (BR-6): детерминированно
+  **красный** re-test / конфликт → прежний `_handle_merge_gate_rollback`. Anti-loop: исчерпание
+  `merge_retest_infra_max_retries` (дефолт 2) → один **инфра-alert** (явно «НЕ дефект кода»,
+  кликабельный номер), задача НЕ уходит в `development`. Kill-switch `merge_retest_infra_tolerance_enabled`
+  off → таймаут = прежний rollback (байт-в-байт).
+- **D4 (контракт необходимости re-test):** при `premerge_rebase_always=True` re-test **пропускается**,
+  когда rebase — доказанный no-op (`merge_gate.head_sha` до==после, обе непусты = ветка уже содержит
+  свежий `origin/main`, тот же коммит уже прошёл CI+tester+staging) → лиз HELD, без re-test. Распространяет
+  на путь `=True` ту же оптимизацию, что путь `=False` уже имеет для не-behind ветки. Fail-safe: любая
+  неопределённость (`head_sha` пуст / git-ошибка) → re-test **выполняется** (BR-6/AC-3 не ослаблен).
+  Kill-switch `merge_retest_skip_when_current_enabled`.
+- **D5 (бюджет):** `merge_retest_timeout_s` 600 → **900** (запас 74% над 516.7s) + валидация
+  `_resolve_retest_timeout` (малформ/непозитив → дефолт 900 + WARNING). Сквозной инвариант ORCH-065/109
+  `reaper_max_running_s (5400) > Σ(deploy-staging gate-work)+grace (≈4460)` соблюдён **без** правки
+  `reaper_max_running_s`.
+- **D6 (наблюдаемость):** in-process счётчики (`_MERGE_GATE_COUNTERS`) + read-only блок `merge_gate` в
+  `GET /queue` (отличим от код-фейл-отката); координация с ORCH-111 (`proc_blocking`) без дубля (ORCH-110
+  предотвращает/толерирует у источника, ORCH-111 наблюдает). Append-only regression-guard:
+  `("ORCH-110", "classify_retest_failure", "src/merge_gate.py")` в `MAIN_REGRESSION_MARKERS`.
+- **Флаги** (`config.py`, дефолт = боевое): `subprocess_tree_kill_enabled`/
+  `merge_retest_infra_tolerance_enabled`/`merge_retest_infra_max_retries`/`merge_retest_infra_retry_delay_s`/
+  `merge_retest_skip_when_current_enabled` (env `ORCH_*`). Откат = выставить 4 kill-switch в `False` +
+  `ORCH_MERGE_RETEST_TIMEOUT_S=600` → байт-в-байт до-ORCH-110. Детали —
+  `docs/work-items/ORCH-110/06-adr/ADR-001-merge-gate-retest-infra-tolerance-and-tree-kill.md`, сквозной
+  `docs/architecture/adr/adr-0042-merge-gate-retest-infra-tolerance-and-tree-kill.md`.
+
 ## Машинный журнал уроков (ORCH-098)
 Шаг 1 («Фундамент», F2) эпика саморазвития: формализует свободнотекстовые «уроки» из `memory/` в
 **машинную структурированную таблицу отклонений конвейера** `lessons`, фундамент для будущих

docs/architecture/README.md

@@ -447,7 +447,10 @@ Self-hosting зацикливался на `deploy-staging`: `scripts/staging_ch
 Назначение: ветка валидируется относительно того `main`, из которого создана; параллельная задача могла уйти вперёд → семантический конфликт слияния (зелёная ветка ломает обновлённый `main`). Merge-gate гарантирует проверку против **актуального** `origin/main` перед слиянием:
 - **Догон:** ветка отстаёт (⇔ `origin/main` не предок HEAD) → `rebase origin/main` в worktree + `push --force-with-lease` (ТОЛЬКО ветка задачи; `main` — никогда). Текстовый конфликт → `rebase --abort` → откат на `development`.
 - **Безусловный pre-merge rebase (ORCH-026, A-2):** при `premerge_rebase_always` (дефолт `True`, скоуп `merge_gate_repos`) short-circuit `branch_is_behind_main` пропускается — `auto_rebase_onto_main` вызывается **всегда** под лизом. На актуальной ветке это no-op (`rebase` не меняет HEAD, `push --force-with-lease` → «Everything up-to-date», CI не триггерится); на отстающей — реальный догон. Детерминированный структурный анти-фантом на уровне планировщика (дополняет рубежи ORCH-073, не заменяет). Kill-switch `premerge_rebase_always=False` → прежнее поведение (ребейз только при behind).
-- **Re-test:** `python -m pytest` (`merge_retest_target`, дефолт `tests/`) в worktree догнанной ветки, тайм-аут `merge_retest_timeout_s`. **Красный** → откат на `development`. **ORCH-110 (design):** (1) re-test исполняется лишь когда rebase реально сдвинул HEAD (`main` уехал); доказанный no-op rebase (ветка уже актуальна, HEAD уже прошёл CI+tester+staging) **пропускает** локальный re-test; (2) **таймаут** re-test — инфра-транзиент (ограниченный повтор + отдельный инфра-alert), а НЕ код-фейл (раньше шёл в откат на `development` + расход developer-retry); (3) спавненный pytest бежит в отдельной группе процессов и при таймауте убивается **деревом** (`os.killpg`) — нет осиротевших процессов (см. ниже).
+- **Re-test:** `python -m pytest` (`merge_retest_target`, дефолт `tests/`) в worktree догнанной ветки, тайм-аут `merge_retest_timeout_s` (ORCH-110 D5: 600 → **900**, валидация `_resolve_retest_timeout` — малформ/непозитив → дефолт 900 + WARNING). **Красный** (детерминированный сбой теста) → откат на `development` + developer-retry (без изменений). **ORCH-110 (реализовано, [adr-0042](adr/adr-0042-merge-gate-retest-infra-tolerance-and-tree-kill.md)):**
+  - **(D4) Контракт необходимости re-test:** локальный re-test исполняется лишь когда rebase реально сдвинул HEAD (`main` уехал). Доказанный **no-op** rebase (`head_sha` до==после, обе непусты — ветка уже содержит свежий `origin/main`, HEAD уже прошёл CI+tester+staging) **пропускает** re-test (`(True, "branch up-to-date (re-test skipped: rebase no-op, HEAD CI-validated)")`, лиз HELD) — распространяет на путь `premerge_rebase_always=True` ту же оптимизацию, что путь `=False` уже имеет для не-behind ветки. Fail-safe: любая неопределённость (`head_sha` пуст / git-ошибка) → re-test **выполняется**. Kill-switch `merge_retest_skip_when_current_enabled`.
+  - **(D2/D3) Толерантность к инфра-таймауту:** предикат `classify_retest_failure(reason)` различает `timeout`/`red`/`lock-busy`/`other`. **Таймаут** re-test = инфра-транзиент → `_handle_merge_gate_infra_retry` (ограниченный повтор по образцу defer: задача **остаётся на `deploy-staging`**, staging-deployer перезапускается с задержкой `merge_retest_infra_retry_delay_s`, **БЕЗ** отката на `development` и **БЕЗ** расхода developer-retry), а НЕ прежний rollback. Исчерпание `merge_retest_infra_max_retries` → один **инфра-alert** (явно «НЕ дефект кода», кликабельный номер), задача НЕ уходит в `development`. Анти-над-толерантность (BR-6): **красный** re-test / конфликт по-прежнему откатывают. Scope-guard: «rebase timeout» (git завис) — НЕ инфра-таймаут re-test, остаётся на rollback-пути. Kill-switch `merge_retest_infra_tolerance_enabled` (off → таймаут=прежний rollback).
+  - **(D1) Tree-kill спавненных pytest:** re-test (и coverage `measure_coverage`) бегут через `proc_group.run_in_process_group` в **отдельной группе процессов** (`start_new_session`); при таймауте убивается **всё дерево** (`os.killpg`, каскад SIGTERM→grace→SIGKILL, грейс = `agent_kill_grace_seconds`) — осиротевшие внуки-pytest не переживают бюджет (корень CPU-голодания инцидента ORCH-109). Контракты возврата сохранены. Fallback never-break: `subprocess_tree_kill_enabled=False`/не-POSIX → прежний `subprocess.run`. Наблюдаемость — блок `merge_gate` в `GET /queue`; координация с ORCH-111 (`proc_blocking`) без дубля.
 - **Сериализация (merge-lock):** файловый **merge-lease** на репо (`<repos_dir>/.merge-lease-<repo>.json`), живёт от гейта до фактического merge. Acquire **неблокирующий** (anti-deadlock при `max_concurrency=1`): busy → **defer** (повторная постановка deployer'а на `deploy-staging` с задержкой через `available_at`), а не откат. Release — на PR-merged вебхуке / `deploy→done` / откате / по возрасту (crash-реклейм). Restart-safe; без изменения схемы БД. **ORCH-026 (A-1):** это окно = «merge → main-updated» (для self `done` ⇔ SHA-in-main, ORCH-073) — пока A не в `main`, B того же репо получает `merge-lock busy` → defer. Окно сериализации per-repo НЕ переписывается; кросс-репо параллелизм сохранён (лиз — per-repo файл).
 - **Условность (как ORCH-35):** реален для `orchestrator`; прочие репо — no-op. Флаги `merge_gate_enabled` / `merge_gate_repos` — поэтапный раскат. Контракт **never-raise**.
 

docs/overview/tech-pipeline.md

@@ -46,7 +46,10 @@ created → analysis → architecture → development → review → testing →
 4. `check_staging_image_fresh` — staging-образ собран из актуального кода.
 
 Это **врезки в переход, а не стадии**: они не появляются в карте `STAGE_TRANSITIONS`, а
-исполняются stage engine'ом внутри ребра. Провал любого из них — откат на доработку. На ребре
+исполняются stage engine'ом внутри ребра. Провал любого из них — откат на доработку. Исключение
+(ORCH-110): **инфра-таймаут** локального re-test merge-gate (а не детерминированный красный тест) —
+это транзиент, а не дефект кода → ограниченный повтор + отдельный инфра-alert, без отката на
+доработку и без расхода developer-retry (красный re-test/конфликт по-прежнему откатывают). На ребре
 `deploy → done` аналогичная врезка merge-verify подтверждает, что код задачи реально слит в
 `main` (слияние — только через PR-API Gitea, см. [интеграции](tech-integrations.md)).
 

src/config.py

@@ -202,18 +202,50 @@ class Settings(BaseSettings):
     #                             only the self-hosting repo (orchestrator). Other
     #                             repos -> conditional no-op (mirrors ORCH-35 staging).
     #   merge_retest_timeout_s -> wall-clock budget for the post-rebase re-test.
+    #       ORCH-110 (D5): raised 600 -> 900 (74% headroom over the observed 516.7s
+    #       suite vs the prior ~16%). Cross-invariant (ORCH-065/109): keep
+    #       reaper_max_running_s (5400) > Σ(deploy-staging gate-work) + grace — see
+    #       docs/work-items/ORCH-110/07-infra-requirements.md.
     #   merge_retest_target    -> pytest target for the re-test (portability across repos).
     #   merge_lock_timeout_s   -> max lease age; an older lease is reclaimed (crash backstop).
     #   merge_defer_delay_s    -> delay before re-running the gate when the lock is busy.
     #   merge_defer_max_attempts -> defer retries before escalation (avoids livelock).
     merge_gate_enabled: bool = True
     merge_gate_repos: str = ""
-    merge_retest_timeout_s: int = 600
+    merge_retest_timeout_s: int = 900
     merge_retest_target: str = "tests/"
     merge_lock_timeout_s: int = 300
     merge_defer_delay_s: int = 60
     merge_defer_max_attempts: int = 5
 
+    # ORCH-110: merge-gate re-test infra-timeout tolerance + tree-kill of the
+    # orchestrator-spawned pytest subprocess (re-test + coverage). Each default =
+    # the desired prod behaviour (ORCH-101 canon); each flag is an INDEPENDENT
+    # kill-switch (off -> byte-for-byte pre-ORCH-110). Detailed ADR:
+    # docs/work-items/ORCH-110/06-adr/ADR-001-merge-gate-retest-infra-tolerance-and-tree-kill.md.
+    #   subprocess_tree_kill_enabled           -> D1: spawn the re-test / coverage
+    #       pytest in its own process group and tree-kill the WHOLE group on timeout
+    #       (no orphan grandchildren grinding the host CPU). off -> the prior
+    #       subprocess.run(timeout=) (ORCH_SUBPROCESS_TREE_KILL_ENABLED).
+    #   merge_retest_infra_tolerance_enabled   -> D3: a re-test TIMEOUT is a transient
+    #       (bounded infra-retry, NOT a code-fault rollback to development burning a
+    #       developer retry). off -> timeout = the prior rollback
+    #       (ORCH_MERGE_RETEST_INFRA_TOLERANCE_ENABLED).
+    #   merge_retest_infra_max_retries         -> D3: infra-retry budget before an
+    #       infra-alert (anti-loop). (ORCH_MERGE_RETEST_INFRA_MAX_RETRIES)
+    #   merge_retest_infra_retry_delay_s       -> D3: delay before the staging-deployer
+    #       re-run. (ORCH_MERGE_RETEST_INFRA_RETRY_DELAY_S)
+    #   merge_retest_skip_when_current_enabled -> D4: skip the local re-test when the
+    #       pre-merge rebase was a PROVEN no-op (branch already at origin/main; HEAD is
+    #       exactly the CI/tester/staging-validated commit). off -> always re-test
+    #       after rebase (ORCH_MERGE_RETEST_SKIP_WHEN_CURRENT_ENABLED).
+    # The tree-kill grace reuses the existing agent_kill_grace_seconds (no new key).
+    subprocess_tree_kill_enabled: bool = True
+    merge_retest_infra_tolerance_enabled: bool = True
+    merge_retest_infra_max_retries: int = 2
+    merge_retest_infra_retry_delay_s: int = 120
+    merge_retest_skip_when_current_enabled: bool = True
+
     # ORCH-036: executable self-deploy (deploy stage drives the host hook).
     # The `deploy` stage for the self-hosting repo is turned into a REAL prod
     # restart via a detached host process, gated by a manual approve. Three-phase

src/coverage_gate.py

@@ -44,7 +44,8 @@ Invariants (ADR-001 §7, never broken):
     artefact / decides. It never calls the deploy hook, never restarts the prod
     container, never pushes / force-pushes ``main``.
 
-This module is a **leaf**: it imports only ``config`` / ``git_worktree`` and lazily
+This module is a **leaf**: it imports only ``config`` / ``git_worktree`` /
+``proc_group`` (the ORCH-110 stdlib-only process-group runner) and lazily
 ``qg.checks.is_self_hosting_repo`` / ``db`` / ``notifications``; it never imports
 ``stage_engine``.
 """
@@ -52,11 +53,11 @@ This module is a **leaf**: it imports only ``config`` / ``git_worktree`` and laz
 import json
 import logging
 import os
-import subprocess
 import sys
 
 from .config import settings
 from .git_worktree import ensure_worktree, get_worktree_path
+from .proc_group import run_in_process_group  # ORCH-110 D1: tree-kill on timeout
 
 logger = logging.getLogger("orchestrator.coverage_gate")
 
@@ -152,22 +153,25 @@ def measure_coverage(repo: str, branch: str) -> float | None:
         "-q",
     ]
     timeout = settings.coverage_run_timeout_s
-    try:
-        subprocess.run(cmd, cwd=wt, capture_output=True, text=True, timeout=timeout)
-    except subprocess.TimeoutExpired:
+    # ORCH-110 (D1 / FR-2 / BR-3): run the coverage suite in its OWN process group so
+    # a timeout tree-kills the WHOLE subtree (the sibling orphan-leak source of the
+    # ORCH-109 incident), not just the direct child. The metric is read from the JSON
+    # file regardless of the exit code, so a non-timeout spawn/OS error (returncode
+    # None) just falls through to "no coverage json produced" -> None, byte-for-byte
+    # the prior fail-open contract.
+    res = run_in_process_group(
+        cmd,
+        cwd=wt,
+        timeout=timeout,
+        tree_kill=bool(getattr(settings, "subprocess_tree_kill_enabled", True)),
+        grace_s=settings.agent_kill_grace_seconds,
+    )
+    if res.timed_out:
         logger.warning(
             "measure_coverage: pytest --cov timed out after %ss for %s/%s",
             timeout, repo, branch,
         )
         return None
-    except FileNotFoundError:
-        logger.warning(
-            "measure_coverage: pytest / pytest-cov not available for %s/%s", repo, branch
-        )
-        return None
-    except (subprocess.SubprocessError, OSError) as e:
-        logger.warning("measure_coverage: pytest --cov error for %s/%s: %s", repo, branch, e)
-        return None
 
     data = None
     try:

src/main.py

@@ -225,6 +225,11 @@ async def queue():
         "reaper": reaper.status(),
         "post_deploy": post_deploy.status(),
         "merge_verify": merge_gate.merge_verify_status(),
+        # ORCH-110 (D6): merge-gate re-test infra-timeout observability (read-only) —
+        # tolerance/skip/tree-kill flags + timeout/infra-retry/skip counters, so an
+        # infra-timeout is distinguishable from a code-fault rollback. Additive block;
+        # never-raise.
+        "merge_gate": merge_gate.merge_gate_status(),
         # ORCH-026 (G-2): declarative task-dependency observability (read-only,
         # NOT a source of truth) — declared edges, blocked tasks, detected cycle.
         "task_deps": task_deps.snapshot(),

src/merge_gate.py

@@ -36,6 +36,7 @@ import time
 
 from .config import settings
 from .git_worktree import ensure_worktree, get_worktree_path
+from .proc_group import run_in_process_group  # ORCH-110 D1: tree-kill on timeout
 
 logger = logging.getLogger("orchestrator.merge_gate")
 
@@ -175,17 +176,103 @@ def auto_rebase_onto_main(repo: str, branch: str) -> tuple[bool, str]:
 # ---------------------------------------------------------------------------
 # re-test in the caught-up worktree
 # ---------------------------------------------------------------------------
+def _resolve_retest_timeout() -> int:
+    """Resolve the re-test wall-clock budget (ORCH-110 D5), never-break.
+
+    ``int(settings.merge_retest_timeout_s)`` when it is a positive int; otherwise
+    the safe default (900) + WARNING (mirrors ``launcher._resolve_timeout``). A
+    malformed / non-positive config never reaches ``subprocess`` (FR-3 / AC-5).
+    """
+    default = 900
+    try:
+        v = int(settings.merge_retest_timeout_s)
+        if v > 0:
+            return v
+        logger.warning(
+            "Non-positive merge_retest_timeout_s=%r; using default %ss",
+            settings.merge_retest_timeout_s, default,
+        )
+    except (TypeError, ValueError):
+        logger.warning(
+            "Invalid merge_retest_timeout_s=%r; using default %ss",
+            settings.merge_retest_timeout_s, default,
+        )
+    return default
+
+
+def head_sha(repo: str, branch: str) -> str:
+    """Return ``git rev-parse HEAD`` of the branch worktree, or ``""`` (ORCH-110 D4).
+
+    Used by ``check_branch_mergeable`` to detect a no-op rebase (HEAD unchanged ->
+    branch already contained the latest origin/main). Reads ONLY the existing
+    per-branch worktree (``get_worktree_path``; the worktree exists by the
+    deploy-staging->deploy edge) — it never CREATES one, so on an absent worktree it
+    safely returns ``""`` (the no-op-skip then never triggers; re-test runs —
+    fail-safe). Never-raise: any git/OS error -> ``""``.
+    """
+    try:
+        wt = get_worktree_path(repo, branch)
+        if not os.path.isdir(wt):
+            return ""
+        r = subprocess.run(
+            ["git", "-C", wt, "rev-parse", "HEAD"],
+            capture_output=True, text=True, timeout=_SHORT_TIMEOUT,
+        )
+        if r.returncode != 0:
+            return ""
+        return (r.stdout or "").strip()
+    except (subprocess.SubprocessError, OSError) as e:
+        logger.warning("head_sha error for %s/%s: %s", repo, branch, e)
+        return ""
+
+
+def classify_retest_failure(reason: str) -> str:
+    """Classify a merge-gate FAIL ``reason`` (ORCH-110 D2). Pure, never-raise.
+
+    The single point for the 'magic string' of the re-test outcome so the routing
+    in ``stage_engine._handle_merge_gate`` can distinguish an INFRA timeout from a
+    real RED re-test without scattering ``"timeout" in reason`` checks:
+
+      * ``"timeout"``   — the re-test hit its wall-clock budget (infra/resource);
+                          ``check_branch_mergeable`` returns ``"re-test timeout after <T>s"``.
+      * ``"red"``       — a deterministically red re-test (a real code defect):
+                          ``"re-test failed after rebase: ..."``.
+      * ``"lock-busy"`` — merge-lock contention (``"merge-lock busy"``).
+      * ``"other"``     — rebase conflict / setup error / anything else.
+
+    SCOPE (ADR D2): ONLY the re-test timeout is an infra-tolerated transient.
+    ``auto_rebase_onto_main``'s own ``"rebase timeout"`` is a DIFFERENT timeout (git
+    hung) and stays on the rollback path — without a successful rebase the branch
+    cannot be caught up to ``main``, so a merge is impossible by construction.
+    """
+    try:
+        r = (reason or "").strip().lower()
+        if r == "merge-lock busy":
+            return "lock-busy"
+        if "re-test timeout" in r:
+            return "timeout"
+        if r.startswith("re-test failed"):
+            return "red"
+        return "other"
+    except Exception:  # noqa: BLE001 - never-raise; unknown -> other (safe: rollback)
+        return "other"
+
+
 def retest_branch(repo: str, branch: str) -> tuple[bool, str]:
     """Run the project test-suite in the (already caught-up) branch worktree.
 
     Command: ``python -m pytest <merge_retest_target>`` (default ``tests/``),
-    matching the orchestrator CI / check_tests_local pattern. Bounded by
-    ``settings.merge_retest_timeout_s``.
+    matching the orchestrator CI / check_tests_local pattern. Bounded by the
+    validated ``_resolve_retest_timeout()`` budget (ORCH-110 D5).
 
-    Returns:
+    ORCH-110 (D1 / FR-2): the suite is spawned in its OWN process group via
+    ``run_in_process_group`` so a timeout tree-kills the WHOLE subtree (no orphan
+    grandchildren grinding the host CPU). The RETURN contract is byte-for-byte the
+    prior one — only the side effect (no leak) changes:
       * ``(True,  "re-test green")``                 — pytest rc == 0
-      * ``(False, "re-test timeout after <T>s")``    — exceeded the timeout (AC-6)
+      * ``(False, "re-test timeout after <T>s")``    — exceeded the timeout (AC-4/AC-6)
       * ``(False, "re-test failed: ...<tail>")``     — non-zero rc, with output tail
+      * ``(False, "re-test error: ...")``            — spawn / OS error (not a timeout)
     Never-raise (AC-9): any setup/OS error -> ``(False, "<reason>")``.
     """
     wt = get_worktree_path(repo, branch)
@@ -197,21 +284,25 @@ def retest_branch(repo: str, branch: str) -> tuple[bool, str]:
             return False, f"re-test setup error: {e}"
 
     target = settings.merge_retest_target or "tests/"
-    timeout = settings.merge_retest_timeout_s
-    try:
-        r = subprocess.run(
-            ["python", "-m", "pytest", target, "-q"],
-            cwd=wt, capture_output=True, text=True, timeout=timeout,
-        )
-    except subprocess.TimeoutExpired:
+    timeout = _resolve_retest_timeout()
+    res = run_in_process_group(
+        ["python", "-m", "pytest", target, "-q"],
+        cwd=wt,
+        timeout=timeout,
+        tree_kill=bool(getattr(settings, "subprocess_tree_kill_enabled", True)),
+        grace_s=settings.agent_kill_grace_seconds,
+    )
+    if res.timed_out:
         logger.warning("retest_branch: timeout (%ss) on %s/%s", timeout, repo, branch)
         return False, f"re-test timeout after {timeout}s"
-    except (subprocess.SubprocessError, OSError) as e:
-        return False, f"re-test error: {e}"
-
-    if r.returncode == 0:
+    if res.returncode is None:
+        # Spawn / OS error that was NOT a timeout -> preserve the prior error reason.
+        detail = (res.stderr or "").strip()[:200]
+        logger.warning("retest_branch: spawn error on %s/%s: %s", repo, branch, detail)
+        return False, f"re-test error: {detail}"
+    if res.returncode == 0:
         return True, "re-test green"
-    tail = ((r.stdout or "") + (r.stderr or ""))[-500:]
+    tail = ((res.stdout or "") + (res.stderr or ""))[-500:]
     logger.warning("retest_branch: red on %s/%s", repo, branch)
     return False, f"re-test failed: ...{tail}"
 
@@ -576,6 +667,82 @@ def merge_verify_status() -> dict:
         return {"enabled": False}
 
 
+# ---------------------------------------------------------------------------
+# ORCH-110 (D6): re-test infra-timeout observability counters (read-only).
+#
+# In-process counters (mirror _MERGE_VERIFY_COUNTERS): reset only on process start,
+# surfaced read-only via merge_gate_status() in GET /queue. NEVER the source of any
+# decision — purely informational, so an infra timeout is distinguishable from a
+# code-fault rollback in the snapshot. never-raise.
+# ---------------------------------------------------------------------------
+_MERGE_GATE_COUNTERS: dict = {
+    "retest_timeout_total": 0,          # re-test hit its wall-clock budget
+    "retest_infra_retry_total": 0,      # infra-timeout -> bounded re-queue (D3)
+    "retest_infra_exhausted_total": 0,  # infra-retry budget exhausted -> infra-alert
+    "retest_skipped_current_total": 0,  # re-test skipped on a no-op rebase (D4)
+    "last_infra_timeout_wi": None,
+}
+
+
+def note_retest_timeout(work_item_id: str | None = None) -> None:
+    """Bump the 're-test timeout' counter (observability only). Never raises."""
+    try:
+        _MERGE_GATE_COUNTERS["retest_timeout_total"] += 1
+        _MERGE_GATE_COUNTERS["last_infra_timeout_wi"] = work_item_id
+    except Exception:  # noqa: BLE001 - observability must never break a decision
+        pass
+
+
+def note_retest_infra_retry() -> None:
+    """Bump the 'infra-timeout re-queue' counter (observability only). Never raises."""
+    try:
+        _MERGE_GATE_COUNTERS["retest_infra_retry_total"] += 1
+    except Exception:  # noqa: BLE001 - observability must never break a decision
+        pass
+
+
+def note_retest_infra_exhausted() -> None:
+    """Bump the 'infra-retry exhausted' counter (observability only). Never raises."""
+    try:
+        _MERGE_GATE_COUNTERS["retest_infra_exhausted_total"] += 1
+    except Exception:  # noqa: BLE001 - observability must never break a decision
+        pass
+
+
+def note_retest_skipped_current() -> None:
+    """Bump the 're-test skipped (no-op rebase)' counter (D4). Never raises."""
+    try:
+        _MERGE_GATE_COUNTERS["retest_skipped_current_total"] += 1
+    except Exception:  # noqa: BLE001 - observability must never break a decision
+        pass
+
+
+def merge_gate_status() -> dict:
+    """Read-only snapshot of the ORCH-110 merge-gate re-test path for GET /queue.
+
+    Additive block; existing /queue keys are untouched. Surfaces the flags +
+    counters so an infra-timeout (retry/exhaustion) is distinguishable from a
+    code-fault rollback. Never the source of any decision. never-raise.
+    """
+    try:
+        return {
+            "infra_tolerance_enabled": bool(settings.merge_retest_infra_tolerance_enabled),
+            "infra_max_retries": settings.merge_retest_infra_max_retries,
+            "infra_retry_delay_s": settings.merge_retest_infra_retry_delay_s,
+            "skip_when_current_enabled": bool(settings.merge_retest_skip_when_current_enabled),
+            "tree_kill_enabled": bool(settings.subprocess_tree_kill_enabled),
+            "retest_timeout_s": settings.merge_retest_timeout_s,
+            "retest_timeout_total": _MERGE_GATE_COUNTERS["retest_timeout_total"],
+            "retest_infra_retry_total": _MERGE_GATE_COUNTERS["retest_infra_retry_total"],
+            "retest_infra_exhausted_total": _MERGE_GATE_COUNTERS["retest_infra_exhausted_total"],
+            "retest_skipped_current_total": _MERGE_GATE_COUNTERS["retest_skipped_current_total"],
+            "last_infra_timeout_wi": _MERGE_GATE_COUNTERS["last_infra_timeout_wi"],
+        }
+    except Exception as e:  # noqa: BLE001 - never-raise contract
+        logger.warning("merge_gate_status error: %s", e)
+        return {"infra_tolerance_enabled": False}
+
+
 def merge_verify_applies(repo: str) -> bool:
     """Whether the ORCH-071 merge-verify under-gate is REAL for this repo.
 
@@ -1037,6 +1204,7 @@ MAIN_REGRESSION_MARKERS: list[tuple[str, str, str]] = [
     ("ORCH-073", "check_main_regression", "src/merge_gate.py"),
     ("ORCH-082", "ensure_open_pr", "src/merge_gate.py"),
     ("ORCH-093", "_classify_merge_response", "src/merge_gate.py"),
+    ("ORCH-110", "classify_retest_failure", "src/merge_gate.py"),
 ]
 
 

src/proc_group.py

@@ -0,0 +1,179 @@
+"""Process-group runner with tree-kill on timeout (ORCH-110 D1).
+
+Background
+----------
+The orchestrator spawns its OWN pytest subprocesses OUTSIDE the LLM-agent path:
+``merge_gate.retest_branch`` (the merge-gate re-test) and
+``coverage_gate.measure_coverage`` (the coverage measurement). Both used a plain
+``subprocess.run(..., timeout=...)``. On ``TimeoutExpired`` CPython kills only the
+DIRECT child (``proc.kill()``); any grandchildren (pytest-xdist workers, tests that
+spawn their own processes) are re-parented to PID 1 and keep running — for days,
+starving the host CPU. That orphan leak is a ROOT cause of the ORCH-109 / PR #129
+incident (a 516s suite blew its 600s budget under CPU starvation -> a false
+merge-gate rollback after a green CI).
+
+This leaf gives those orchestrator-spawned runs the SAME SIGTERM->grace->SIGKILL
+cascade the LLM agents already get from ``launcher.stop_process`` — but at the
+process-GROUP level so the WHOLE subtree dies on a timeout:
+
+  * the child becomes the leader of a NEW session / process group
+    (``start_new_session=True`` -> ``setsid``), so every descendant shares
+    ``pgid == child.pid``;
+  * on timeout the whole group is killed with ``os.killpg`` (SIGTERM, grace poll,
+    then SIGKILL), then reaped (no zombies).
+
+Invariants:
+  * **stdlib-only leaf** — imports only ``os`` / ``signal`` / ``subprocess`` /
+    ``time`` / ``logging`` / ``dataclasses``; NEVER another ``src/*`` module
+    (mirrors the purity of ``serial_gate`` / ``staging_verdict``).
+  * **never-raise** — any OS error degrades to a safe ``ProcResult`` (the caller maps
+    it to its existing contract); an exception never escapes.
+  * **never-break fallback** — with ``tree_kill=False`` OR on a platform without
+    ``os.killpg`` / ``os.setsid`` (non-POSIX) it degrades to the prior
+    ``subprocess.run(cmd, ..., timeout=timeout)`` (byte-for-byte the old behaviour;
+    prod is Linux and never hits the fallback).
+"""
+
+import logging
+import os
+import signal
+import subprocess
+import time
+from dataclasses import dataclass
+
+logger = logging.getLogger("orchestrator.proc_group")
+
+
+@dataclass
+class ProcResult:
+    """Outcome of a process-group run (ORCH-110 D1).
+
+    ``returncode`` is the child's exit code, or ``None`` when the process could not
+    be reaped with a real code (timeout-kill / spawn error). ``timed_out`` is True
+    iff the command exceeded its wall-clock budget (and the whole group was killed).
+    """
+
+    returncode: int | None
+    stdout: str
+    stderr: str
+    timed_out: bool
+
+
+def _tree_kill_supported() -> bool:
+    """POSIX process-group ops available? (non-POSIX -> fall back to subprocess.run)."""
+    return hasattr(os, "killpg") and hasattr(os, "getpgid") and hasattr(os, "setsid")
+
+
+def _kill_group(pid: int, grace_s: float) -> None:
+    """SIGTERM -> grace -> SIGKILL the whole process GROUP led by ``pid``.
+
+    Mirrors ``launcher.stop_process`` but targets the process GROUP (``os.killpg``)
+    so grandchildren die too. ``ProcessLookupError`` is tolerated at every step (the
+    group may already be gone). Never raises.
+    """
+    try:
+        pgid = os.getpgid(pid)
+    except (ProcessLookupError, OSError):
+        return
+
+    # Phase 1: SIGTERM the whole group (graceful).
+    try:
+        os.killpg(pgid, signal.SIGTERM)
+    except ProcessLookupError:
+        return
+    except OSError as e:  # noqa: BLE001 - never-raise
+        logger.warning("proc_group: SIGTERM killpg(%s) error: %s", pgid, e)
+
+    # Phase 2: poll for graceful group exit within the grace window.
+    poll = 0.2
+    waited = 0.0
+    while waited < max(0.0, grace_s):
+        time.sleep(poll)
+        waited += poll
+        try:
+            os.killpg(pgid, 0)  # signal 0 = liveness probe of the group
+        except ProcessLookupError:
+            return  # whole group gone
+        except OSError:
+            break
+
+    # Phase 3: still alive -> hard SIGKILL the group.
+    try:
+        os.killpg(pgid, signal.SIGKILL)
+    except ProcessLookupError:
+        return
+    except OSError as e:  # noqa: BLE001 - never-raise
+        logger.warning("proc_group: SIGKILL killpg(%s) error: %s", pgid, e)
+
+
+def _run_plain(cmd, *, cwd, timeout, env=None) -> ProcResult:
+    """Fallback path: prior ``subprocess.run`` semantics (never-break)."""
+    try:
+        r = subprocess.run(
+            cmd, cwd=cwd, env=env, capture_output=True, text=True, timeout=timeout
+        )
+        return ProcResult(
+            returncode=r.returncode, stdout=r.stdout or "", stderr=r.stderr or "",
+            timed_out=False,
+        )
+    except subprocess.TimeoutExpired:
+        return ProcResult(returncode=None, stdout="", stderr="", timed_out=True)
+    except (subprocess.SubprocessError, OSError) as e:
+        logger.warning("proc_group(_run_plain): error for %s: %s", cmd, e)
+        return ProcResult(returncode=None, stdout="", stderr=str(e), timed_out=False)
+
+
+def run_in_process_group(
+    cmd, *, cwd, timeout, env=None, grace_s: float = 5.0, tree_kill: bool = True
+) -> ProcResult:
+    """Run ``cmd`` in its own process group; tree-kill the whole group on timeout.
+
+    See the module docstring. Returns a ``ProcResult``; ``timed_out=True`` iff the
+    command exceeded ``timeout`` (and the whole group was killed). Never raises.
+
+    Fallback (never-break): ``tree_kill=False`` or a non-POSIX platform -> a plain
+    ``subprocess.run(cmd, ..., timeout=timeout)`` (the prior behaviour).
+    """
+    if not tree_kill or not _tree_kill_supported():
+        return _run_plain(cmd, cwd=cwd, timeout=timeout, env=env)
+
+    try:
+        proc = subprocess.Popen(  # noqa: S603 - cmd is an internal, fixed argv
+            cmd,
+            cwd=cwd,
+            env=env,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
+            text=True,
+            start_new_session=True,  # setsid -> child leads a new process group
+        )
+    except (OSError, ValueError) as e:
+        logger.warning("proc_group: spawn error for %s: %s", cmd, e)
+        return ProcResult(returncode=None, stdout="", stderr=str(e), timed_out=False)
+
+    try:
+        out, err = proc.communicate(timeout=timeout)
+        return ProcResult(
+            returncode=proc.returncode, stdout=out or "", stderr=err or "",
+            timed_out=False,
+        )
+    except subprocess.TimeoutExpired:
+        # Tree-kill the WHOLE group (children + grandchildren), then reap to avoid
+        # leaving a zombie behind.
+        _kill_group(proc.pid, grace_s)
+        out, err = "", ""
+        try:
+            out, err = proc.communicate(timeout=grace_s + 5.0)
+        except (subprocess.TimeoutExpired, OSError, ValueError):
+            try:
+                proc.wait(timeout=grace_s + 5.0)
+            except (subprocess.TimeoutExpired, OSError, ValueError):
+                pass
+        return ProcResult(
+            returncode=proc.returncode, stdout=out or "", stderr=err or "",
+            timed_out=True,
+        )
+    except (OSError, ValueError) as e:  # noqa: BLE001 - never-raise
+        logger.warning("proc_group: communicate error for %s: %s", cmd, e)
+        _kill_group(proc.pid, grace_s)
+        return ProcResult(returncode=proc.returncode, stdout="", stderr=str(e), timed_out=False)

src/qg/checks.py

@@ -708,10 +708,38 @@ def check_branch_mergeable(repo: str, work_item_id: str, branch: str) -> tuple[b
                 logger.info("check_branch_mergeable: %s up-to-date with main", branch)
                 return True, "branch up-to-date with main"
 
+            # ORCH-110 (D4): capture HEAD before/after the rebase to detect a no-op
+            # rebase (branch already contained the latest origin/main).
+            pre_sha = merge_gate.head_sha(repo, branch)
             ok, rb_reason = merge_gate.auto_rebase_onto_main(repo, branch)
             if not ok:
                 merge_gate.release_merge_lease(repo, branch)
                 return False, rb_reason  # "rebase conflict: ..."
+            post_sha = merge_gate.head_sha(repo, branch)
+
+            # ORCH-110 (D4 / FR-4 / AC-6): re-test contract. ORCH-043 catches a
+            # SEMANTIC merge conflict that can only arise when ``main`` actually moved
+            # and the branch was really rebased onto new commits. When the rebase was
+            # a PROVEN no-op (HEAD unchanged), there is no "moved main" -> the local
+            # re-test re-checks exactly the commit CI + tester + staging already
+            # validated on THIS HEAD -> it is a redundant single point of false
+            # failure (the ORCH-109 timeout incident). Skip it ONLY on a proven no-op
+            # (both SHAs non-empty AND equal); on ANY uncertainty (empty SHA / flag
+            # off) the re-test runs (fail-safe to BR-6/AC-3). This extends to the
+            # premerge_rebase_always=True path the same optimisation the
+            # premerge_rebase_always=False not-behind short-circuit already has.
+            if (
+                bool(getattr(settings, "merge_retest_skip_when_current_enabled", False))
+                and pre_sha and post_sha and pre_sha == post_sha
+            ):
+                logger.info(
+                    "check_branch_mergeable: %s rebase no-op (HEAD %s unchanged) -> "
+                    "re-test skipped (HEAD CI-validated)", branch, pre_sha[:8],
+                )
+                merge_gate.note_retest_skipped_current()
+                return True, (
+                    "branch up-to-date (re-test skipped: rebase no-op, HEAD CI-validated)"
+                )
 
             ok_t, t_reason = merge_gate.retest_branch(repo, branch)
             if ok_t:

src/stage_engine.py

@@ -995,6 +995,22 @@ def _handle_merge_gate(
         )
         return True
 
+    # ORCH-110 (D2/D3): classify the FAIL cause. An INFRA re-test TIMEOUT (the
+    # ORCH-109 incident: green CI + green tester, but the local re-test blew its
+    # wall-clock budget under CPU starvation) is a TRANSIENT, NOT a code fault — it
+    # must NOT take the _handle_merge_gate_rollback path (rollback to development +
+    # burn a developer retry that nobody can fix). Route it to a bounded infra-retry
+    # instead. A deterministically RED re-test / conflict still rolls back (BR-6 /
+    # AC-3). Kill-switch off -> byte-for-byte the prior behaviour (timeout ->
+    # rollback), NFR-2.
+    if merge_gate.classify_retest_failure(reason) == "timeout":
+        merge_gate.note_retest_timeout(work_item_id)
+        if settings.merge_retest_infra_tolerance_enabled:
+            _handle_merge_gate_infra_retry(
+                task_id, current_stage, repo, work_item_id, branch, reason, result
+            )
+            return True
+
     _handle_merge_gate_rollback(
         task_id, current_stage, repo, work_item_id, branch, reason, result
     )
@@ -1044,6 +1060,110 @@ def _handle_merge_gate_defer(
         )
 
 
+def _merge_infra_retry_count(task_id: int) -> int:
+    """How many times this task was re-queued by the merge-gate infra-timeout path.
+
+    Restart-safe (mirror of ``_merge_defer_count``): counted from the persisted jobs
+    queue by the infra-retry marker in task_content, so a service restart never
+    resets the infra-retry budget (ORCH-110 D3 / AC-9).
+    """
+    conn = get_db()
+    n = conn.execute(
+        "SELECT COUNT(*) FROM jobs WHERE task_id=? AND "
+        "task_content LIKE '%merge-gate infra-timeout retry%'",
+        (task_id,),
+    ).fetchone()[0]
+    conn.close()
+    return n
+
+
+def _handle_merge_gate_infra_retry(
+    task_id, current_stage, repo, work_item_id, branch, reason, result: AdvanceResult
+):
+    """ORCH-110 (D3): merge-gate re-test INFRA timeout -> bounded infra-retry.
+
+    Mirror of ``_handle_merge_gate_defer`` (NOT ``_handle_merge_gate_rollback``): the
+    task STAYS on deploy-staging, the staging-deployer is re-queued after a delay, and
+    there is NO ``update_task_stage("development")``, NO developer-retry increment and
+    NO code-fault QG-failure semantics — so a green-CI + green-tester task is never
+    falsely sent back to development for an infra hiccup (BR-1 / AC-1). The merge lease
+    was already released by ``check_branch_mergeable`` on the timeout, so the re-run
+    re-acquires it normally (defer-path parity).
+
+    Bounded (anti-loop, NFR-5 / AC-9): after ``merge_retest_infra_max_retries``
+    re-tries the task is blocked with an INFRA-specific alert (a resource problem,
+    explicitly NOT "developer must fix"), so it never bounces forever. The first
+    timeout already tree-killed the orphan processes (D1), so the next re-test usually
+    passes — the retry is recovery, not masking. The task does NOT move to development.
+
+    never-raise (TC-09): any unexpected error in this transient path is logged with a
+    WARNING and swallowed (the task simply stays on deploy-staging for the reconciler /
+    reaper to re-drive) so an exception never escapes into ``advance_stage``.
+    """
+    try:
+        _merge_gate_infra_retry_impl(
+            task_id, current_stage, repo, work_item_id, branch, reason, result
+        )
+    except Exception as e:  # noqa: BLE001 - never-raise; do not crash advance_stage
+        logger.warning(
+            f"Task {task_id}: merge-gate infra-retry path error: {e} "
+            f"(swallowed; task stays on deploy-staging)"
+        )
+        result.note = "merge-gate-infra-retry-error"
+
+
+def _merge_gate_infra_retry_impl(
+    task_id, current_stage, repo, work_item_id, branch, reason, result: AdvanceResult
+):
+    """Body of the merge-gate infra-retry (ORCH-110 D3); wrapped by the never-raise
+    ``_handle_merge_gate_infra_retry``."""
+    retries = _merge_infra_retry_count(task_id)
+    if retries < settings.merge_retest_infra_max_retries:
+        task_desc = (
+            f"Work item: {work_item_id}\nRepo: {repo}\nBranch: {branch}\n"
+            f"Stage: deploy-staging\nNote: merge-gate infra-timeout retry "
+            f"(attempt {retries + 1}/{settings.merge_retest_infra_max_retries}) — "
+            f"re-test hit the wall-clock budget (infra/resource, NOT a code defect); "
+            f"retrying after {settings.merge_retest_infra_retry_delay_s}s. Reason: {reason}."
+        )
+        new_job = enqueue_job(
+            "deployer", repo, task_desc, task_id=task_id,
+            available_at_delay_s=settings.merge_retest_infra_retry_delay_s,
+        )
+        merge_gate.note_retest_infra_retry()
+        result.enqueued_agent = "deployer"
+        result.enqueued_job_id = new_job
+        result.note = "merge-gate-infra-retry"
+        logger.warning(
+            f"Task {task_id}: merge-gate re-test infra-timeout, re-queued deployer "
+            f"(job_id={new_job}, attempt {retries + 1}/"
+            f"{settings.merge_retest_infra_max_retries}); NOT a code fault ({reason})"
+        )
+    else:
+        merge_gate.note_retest_infra_exhausted()
+        set_issue_blocked(work_item_id)
+        send_telegram(
+            f"⏳ {link_for(work_item_id)}: merge-gate re-test infra-timeout "
+            f"сохраняется после {settings.merge_retest_infra_max_retries} повторов — "
+            f"ресурсная проблема (CPU / осиротевшие процессы), НЕ дефект кода. "
+            f"Проверьте хост / нужно ручное вмешательство."
+        )
+        plane_add_comment(
+            work_item_id,
+            f"⏳ Merge-gate re-test infra-timeout сохраняется после "
+            f"{settings.merge_retest_infra_max_retries} повторов — инфраструктурная "
+            f"проблема (ресурсы / осиротевшие процессы), НЕ дефект кода. Задача "
+            f"остаётся на deploy-staging; нужно ручное вмешательство (НЕ доработка кода).",
+            author="deployer",
+        )
+        result.alerted = True
+        result.note = "merge-gate-infra-retry-exhausted"
+        logger.error(
+            f"Task {task_id}: merge-gate re-test infra-timeout exhausted "
+            f"({settings.merge_retest_infra_max_retries} retries) ({reason})"
+        )
+
+
 def _handle_merge_gate_rollback(
     task_id, current_stage, repo, work_item_id, branch, reason, result: AdvanceResult
 ):

tests/test_config.py

@@ -56,7 +56,7 @@ def test_merge_gate_settings_defaults(monkeypatch):
     s = Settings()
     assert s.merge_gate_enabled is True
     assert s.merge_gate_repos == ""
-    assert s.merge_retest_timeout_s == 600
+    assert s.merge_retest_timeout_s == 900  # ORCH-110 D5: raised 600 -> 900
     assert s.merge_retest_target == "tests/"
     assert s.merge_lock_timeout_s == 300
     assert s.merge_defer_delay_s == 60

tests/test_coverage_gate.py

@@ -425,12 +425,14 @@ def test_tc14_real_measurement(tmp_path, monkeypatch):
 
 
 def test_tc14_measure_timeout_returns_none(monkeypatch):
-    import subprocess
+    # ORCH-110: measure_coverage now runs via proc_group.run_in_process_group
+    # (tree-kill on timeout). A timed_out ProcResult -> None (prior contract).
+    from src.proc_group import ProcResult
     monkeypatch.setattr(cg, "ensure_worktree", lambda r, b: "/tmp")
-
-    def _timeout(*a, **k):
-        raise subprocess.TimeoutExpired(cmd="pytest", timeout=1)
-    monkeypatch.setattr(cg.subprocess, "run", _timeout)
+    monkeypatch.setattr(
+        cg, "run_in_process_group",
+        lambda *a, **k: ProcResult(returncode=None, stdout="", stderr="", timed_out=True),
+    )
     assert cg.measure_coverage(_REPO, _BRANCH) is None
 
 

tests/test_merge_gate.py

@@ -203,10 +203,16 @@ def fake_worktree(tmp_path, monkeypatch):
     return str(wt)
 
 
+# ORCH-110: retest_branch now runs the suite via proc_group.run_in_process_group
+# (tree-kill on timeout). The seam is mocked with a ProcResult; the RETURN contract
+# is byte-for-byte the prior one.
+from src.proc_group import ProcResult  # noqa: E402
+
+
 def test_tc07_retest_green(fake_worktree, monkeypatch):
     monkeypatch.setattr(
-        merge_gate.subprocess, "run",
-        lambda *a, **k: subprocess.CompletedProcess(a, 0, "1 passed", ""),
+        merge_gate, "run_in_process_group",
+        lambda *a, **k: ProcResult(returncode=0, stdout="1 passed", stderr="", timed_out=False),
     )
     ok, reason = merge_gate.retest_branch("orchestrator", "feature/x")
     assert ok is True
@@ -215,9 +221,11 @@ def test_tc07_retest_green(fake_worktree, monkeypatch):
 
 def test_tc08_retest_red_with_tail(fake_worktree, monkeypatch):
     monkeypatch.setattr(
-        merge_gate.subprocess, "run",
-        lambda *a, **k: subprocess.CompletedProcess(
-            a, 1, "FAILED tests/test_x.py::t - AssertionError\n1 failed", ""
+        merge_gate, "run_in_process_group",
+        lambda *a, **k: ProcResult(
+            returncode=1,
+            stdout="FAILED tests/test_x.py::t - AssertionError\n1 failed",
+            stderr="", timed_out=False,
         ),
     )
     ok, reason = merge_gate.retest_branch("orchestrator", "feature/x")
@@ -227,11 +235,11 @@ def test_tc08_retest_red_with_tail(fake_worktree, monkeypatch):
 
 
 def test_tc09_retest_timeout(fake_worktree, monkeypatch):
-    def _boom(*a, **k):
-        raise subprocess.TimeoutExpired(cmd="pytest", timeout=1)
-
     monkeypatch.setattr(merge_gate.settings, "merge_retest_timeout_s", 1)
-    monkeypatch.setattr(merge_gate.subprocess, "run", _boom)
+    monkeypatch.setattr(
+        merge_gate, "run_in_process_group",
+        lambda *a, **k: ProcResult(returncode=None, stdout="", stderr="", timed_out=True),
+    )
     ok, reason = merge_gate.retest_branch("orchestrator", "feature/x")
     assert ok is False
     assert "re-test timeout" in reason

tests/test_orch110_budget_invariants.py

@@ -0,0 +1,82 @@
+"""ORCH-110 TC-08: re-test budget validation + cross-invariants (D5).
+
+Covers FR-3 / AC-5 / NFR-6:
+  * ``_resolve_retest_timeout`` validates the config (malformed / non-positive ->
+    safe default 900 + WARNING; never reaches subprocess);
+  * the budget was bumped 600 -> 900;
+  * the cross-invariant ``reaper_max_running_s > Σ(deploy-staging gate-work) + grace``
+    (ORCH-065/109) still holds with the new 900s re-test budget — WITHOUT raising
+    ``reaper_max_running_s``.
+"""
+import os
+import tempfile
+
+os.environ.setdefault("ORCH_DB_PATH", os.path.join(tempfile.gettempdir(), "test_orch110_budget.db"))
+os.environ.setdefault("ORCH_GITEA_TOKEN", "test-token")
+os.environ.setdefault("ORCH_PLANE_API_TOKEN", "test-token")
+
+import pytest  # noqa: E402
+
+from src import merge_gate  # noqa: E402
+from src.config import Settings, settings  # noqa: E402
+
+
+# ---------------------------------------------------------------------------
+# _resolve_retest_timeout — validation (never-break).
+# ---------------------------------------------------------------------------
+def test_tc08_resolve_uses_positive_config(monkeypatch):
+    monkeypatch.setattr(merge_gate.settings, "merge_retest_timeout_s", 1234, raising=False)
+    assert merge_gate._resolve_retest_timeout() == 1234
+
+
+@pytest.mark.parametrize("bad", [0, -5, "abc", None, 3.0])
+def test_tc08_resolve_bad_config_falls_back_to_default(monkeypatch, bad):
+    monkeypatch.setattr(merge_gate.settings, "merge_retest_timeout_s", bad, raising=False)
+    # 3.0 is a valid positive int(3) -> stays 3; everything else -> 900 default.
+    out = merge_gate._resolve_retest_timeout()
+    if bad == 3.0:
+        assert out == 3
+    else:
+        assert out == 900
+
+
+def test_tc08_default_budget_bumped_to_900():
+    """D5: the shipped default budget is 900 (raised from 600)."""
+    assert Settings().merge_retest_timeout_s == 900
+
+
+# ---------------------------------------------------------------------------
+# Cross-invariant: reaper backstop covers the worst-case deploy-staging gate-work.
+# ---------------------------------------------------------------------------
+def test_tc08_reaper_covers_deploy_staging_worstcase():
+    """ORCH-065/109 invariant with the new 900s re-test budget (ADR D5 table)."""
+    s = Settings()
+    # Worst-case sum of work charged to a deploy-staging-deployer job (ADR D5).
+    security = 120
+    rebase = 120
+    image = 600
+    worst = (
+        s.agent_timeout_seconds          # deployer agent (1800)
+        + security
+        + rebase
+        + s.merge_retest_timeout_s       # re-test (900, new)
+        + s.coverage_run_timeout_s       # coverage (900)
+        + image
+        + s.agent_kill_grace_seconds     # grace (20)
+    )
+    assert worst <= 4460  # matches the ADR D5 table
+    assert s.reaper_max_running_s > worst, (
+        f"reaper_max_running_s={s.reaper_max_running_s} must exceed "
+        f"deploy-staging worst-case {worst}"
+    )
+
+
+def test_tc08_reaper_still_covers_max_agent_timeout():
+    """ORCH-065/109: reaper_max_running_s > max(agent timeout) + grace (unchanged)."""
+    s = Settings()
+    assert s.reaper_max_running_s > s.agent_timeout_developer_s + s.agent_kill_grace_seconds
+
+
+def test_tc08_reaper_max_running_s_unchanged():
+    """D5 must NOT change reaper_max_running_s (stays 5400 from ORCH-109)."""
+    assert settings.reaper_max_running_s == 5400

tests/test_orch110_false_rollback_regression.py

@@ -0,0 +1,213 @@
+"""ORCH-110 TC-10: regression of the ORCH-109 / PR #129 incident.
+
+Incident: tester PASS + green CI + the branch not-behind, but the merge-gate local
+re-test blew its wall-clock budget under CPU starvation -> ``check_branch_mergeable``
+returned ``(False, "re-test timeout ...")`` -> the engine routed it to
+``_handle_merge_gate_rollback`` (rollback deploy-staging -> development + a developer
+retry) -> every retry timed out the same way -> "Merge-gate still failing after 3
+developer retries" -> a stuck task needing manual intervention.
+
+This drives the REAL ``check_branch_mergeable`` through ``advance_stage`` (only the
+git/test primitives are mocked) and asserts the incident can no longer happen:
+
+  * Scenario A (D4) — a not-behind branch (no-op rebase) on a green-CI HEAD: the
+    local re-test is SKIPPED entirely -> the gate passes -> advance to deploy.
+  * Scenario B (D3) — a real catch-up whose re-test times out: a bounded infra-retry
+    (task stays on deploy-staging), NEVER a rollback to development and NEVER the
+    "Merge-gate still failing after N developer retries" alert.
+
+RED-before / GREEN-after: on pre-ORCH-110 code both scenarios would roll the task
+back to development (Scenario A would even run the doomed re-test), so the
+``rolled_back_to is None`` / ``stage == deploy-staging`` assertions below would FAIL.
+"""
+import os
+import tempfile
+
+import pytest
+
+_test_db = os.path.join(tempfile.gettempdir(), "test_orch110_regression.db")
+os.environ["ORCH_DB_PATH"] = _test_db
+os.environ["ORCH_REPOS_DIR"] = tempfile.gettempdir()
+os.environ.setdefault("ORCH_GITEA_TOKEN", "test-token")
+os.environ.setdefault("ORCH_PLANE_API_TOKEN", "test-token")
+
+from unittest.mock import MagicMock  # noqa: E402
+
+import src.db as _db  # noqa: E402
+from src.db import init_db, get_db  # noqa: E402
+from src import stage_engine  # noqa: E402
+from src import merge_gate  # noqa: E402
+from src.qg import checks as qg  # noqa: E402
+from src.stage_engine import advance_stage  # noqa: E402
+
+_REPO = "orchestrator"
+_WI = "ORCH-110"
+_BRANCH = "feature/ORCH-110-x"
+
+
+@pytest.fixture(autouse=True)
+def fresh_db(monkeypatch):
+    monkeypatch.setattr(_db.settings, "db_path", _test_db)
+    if os.path.exists(_test_db):
+        os.unlink(_test_db)
+    init_db()
+    yield
+
+
+@pytest.fixture(autouse=True)
+def silence(monkeypatch):
+    for name in (
+        "notify_stage_change", "notify_qg_failure", "send_telegram",
+        "plane_notify_stage", "plane_notify_qg", "plane_add_comment",
+        "set_issue_in_progress", "set_issue_blocked", "notify_approve_requested",
+        "set_issue_in_review", "set_issue_needs_input",
+    ):
+        monkeypatch.setattr(stage_engine, name, MagicMock())
+    # Keep the merge-gate the only intervening sub-gate on the edge; no Phase A.
+    monkeypatch.setattr(stage_engine.settings, "deploy_require_manual_approve", False)
+    # Real gate scope: orchestrator self-hosting.
+    monkeypatch.setattr(qg.settings, "merge_gate_enabled", True, raising=False)
+    monkeypatch.setattr(qg.settings, "merge_gate_repos", "", raising=False)
+    monkeypatch.setattr(qg.settings, "premerge_rebase_always", True, raising=False)
+    monkeypatch.setattr(qg.settings, "merge_retest_skip_when_current_enabled", True, raising=False)
+    monkeypatch.setattr(stage_engine.settings, "merge_retest_infra_tolerance_enabled", True)
+    monkeypatch.setattr(stage_engine.settings, "merge_retest_infra_max_retries", 2)
+    monkeypatch.setattr(stage_engine.settings, "merge_retest_infra_retry_delay_s", 120)
+
+
+def _pass(*a, **k):
+    return (True, "ok")
+
+
+def _patch_edge_gates_except_merge(monkeypatch):
+    """All edge gates pass EXCEPT check_branch_mergeable, which stays the REAL one."""
+    patched = {**stage_engine.QG_CHECKS}
+    patched["check_staging_status"] = _pass
+    patched["check_security_gate"] = _pass
+    patched["check_coverage_gate"] = _pass
+    patched["check_staging_image_fresh"] = _pass
+    monkeypatch.setattr(stage_engine, "QG_CHECKS", patched)
+
+
+def _mock_merge_primitives(monkeypatch, *, head_shas, retest_result, retest_calls):
+    """Mock the git/test primitives the real check_branch_mergeable composes."""
+    monkeypatch.setattr(merge_gate, "acquire_merge_lease", lambda *a, **k: (True, "lease acquired"))
+    monkeypatch.setattr(merge_gate, "release_merge_lease", lambda *a, **k: None)
+    monkeypatch.setattr(merge_gate, "branch_is_behind_main", lambda r, b: True)
+    monkeypatch.setattr(merge_gate, "auto_rebase_onto_main", lambda r, b: (True, "rebased onto origin/main"))
+
+    shas = list(head_shas)
+
+    def _head_sha(r, b):
+        return shas.pop(0) if shas else ""
+
+    monkeypatch.setattr(merge_gate, "head_sha", _head_sha)
+
+    def _retest(r, b):
+        retest_calls.append((r, b))
+        return retest_result
+
+    monkeypatch.setattr(merge_gate, "retest_branch", _retest)
+
+
+def _make_task():
+    conn = get_db()
+    cur = conn.execute(
+        "INSERT INTO tasks (plane_id, work_item_id, repo, branch, stage) VALUES (?,?,?,?,?)",
+        (f"plane-{_WI}", _WI, _REPO, _BRANCH, "deploy-staging"),
+    )
+    tid = cur.lastrowid
+    conn.commit()
+    conn.close()
+    return tid
+
+
+def _stage(task_id):
+    conn = get_db()
+    row = conn.execute("SELECT stage FROM tasks WHERE id=?", (task_id,)).fetchone()
+    conn.close()
+    return row[0]
+
+
+def _agents():
+    conn = get_db()
+    rows = conn.execute("SELECT agent FROM jobs ORDER BY id").fetchall()
+    conn.close()
+    return [r[0] for r in rows]
+
+
+def _advance(task_id):
+    return advance_stage(task_id, "deploy-staging", _REPO, _WI, _BRANCH, finished_agent="deployer")
+
+
+# ---------------------------------------------------------------------------
+# Scenario A (D4): the literal incident — not-behind branch, green CI HEAD.
+# The re-test is SKIPPED, so the timeout can never happen; advance to deploy.
+# ---------------------------------------------------------------------------
+def test_tc10_incident_noop_rebase_skips_retest_and_advances(monkeypatch):
+    _patch_edge_gates_except_merge(monkeypatch)
+    retest_calls = []
+    # Same SHA before/after rebase -> proven no-op. retest is rigged to TIME OUT if
+    # it were (wrongly) called — proving the skip is what avoids the false rollback.
+    _mock_merge_primitives(
+        monkeypatch,
+        head_shas=["abc123", "abc123"],
+        retest_result=(False, "re-test timeout after 900s"),
+        retest_calls=retest_calls,
+    )
+    task_id = _make_task()
+    res = _advance(task_id)
+
+    assert retest_calls == [], "re-test must be SKIPPED on a no-op rebase (D4)"
+    assert res.rolled_back_to is None              # the incident's false rollback is gone
+    assert res.advanced is True
+    assert res.to_stage == "deploy"
+    assert _stage(task_id) == "deploy"
+    assert "developer" not in _agents()           # no developer-retry burned
+    assert stage_engine.set_issue_blocked.called is False
+
+
+# ---------------------------------------------------------------------------
+# Scenario B (D3): a real catch-up whose re-test times out -> infra-retry, not
+# rollback, and never the "still failing after N retries" manual gate.
+# ---------------------------------------------------------------------------
+def test_tc10_real_catchup_retest_timeout_infra_retries_not_rollback(monkeypatch):
+    _patch_edge_gates_except_merge(monkeypatch)
+    retest_calls = []
+    # HEAD moved (real catch-up) -> re-test runs -> times out.
+    _mock_merge_primitives(
+        monkeypatch,
+        head_shas=["old111", "new222"],
+        retest_result=(False, "re-test timeout after 900s"),
+        retest_calls=retest_calls,
+    )
+    task_id = _make_task()
+    res = _advance(task_id)
+
+    assert retest_calls, "re-test SHOULD run on a real (HEAD-moving) catch-up"
+    assert res.rolled_back_to is None              # NOT the code-fault rollback
+    assert res.note == "merge-gate-infra-retry"
+    assert _stage(task_id) == "deploy-staging"     # stays put for a bounded retry
+    assert _agents() == ["deployer"]               # staging-deployer re-queued, NOT developer
+    assert stage_engine.set_issue_blocked.called is False
+    # The "Merge-gate still failing after N developer retries" manual gate never fires.
+    for call in stage_engine.send_telegram.call_args_list:
+        assert "developer retries" not in call[0][0]
+
+
+def test_tc10_real_catchup_red_retest_still_rolls_back(monkeypatch):
+    """Anti-over-tolerance guard inside the incident scenario: a genuinely RED
+    re-test on a real catch-up STILL rolls back (BR-6 / AC-3)."""
+    _patch_edge_gates_except_merge(monkeypatch)
+    retest_calls = []
+    _mock_merge_primitives(
+        monkeypatch,
+        head_shas=["old111", "new222"],
+        retest_result=(False, "re-test failed: ...1 failed"),
+        retest_calls=retest_calls,
+    )
+    task_id = _make_task()
+    res = _advance(task_id)
+    assert res.rolled_back_to == "development"
+    assert _stage(task_id) == "development"
+    assert _agents() == ["developer"]

tests/test_orch110_killswitch.py

@@ -0,0 +1,103 @@
+"""ORCH-110 TC-07: kill-switches + non-self repo => byte-for-byte pre-ORCH-110.
+
+Covers NFR-2 / FR-5 / AC-7. Each ORCH-110 behaviour is an INDEPENDENT kill-switch;
+with a flag off the affected path reverts to the prior behaviour. A non-self-hosting
+repo (enduro-trails) never reaches the infra-retry / tree-kill paths at all (the
+merge-gate is N/A for it).
+"""
+import os
+import tempfile
+
+os.environ.setdefault("ORCH_DB_PATH", os.path.join(tempfile.gettempdir(), "test_orch110_killswitch.db"))
+os.environ.setdefault("ORCH_REPOS_DIR", tempfile.gettempdir())
+os.environ.setdefault("ORCH_GITEA_TOKEN", "test-token")
+os.environ.setdefault("ORCH_PLANE_API_TOKEN", "test-token")
+
+import pytest  # noqa: E402
+
+from src import merge_gate  # noqa: E402
+from src.qg import checks as qg  # noqa: E402
+from src.proc_group import ProcResult  # noqa: E402
+
+_REPO = "orchestrator"
+_BRANCH = "feature/ORCH-110-x"
+_WI = "ORCH-110"
+
+
+# ---------------------------------------------------------------------------
+# D1 kill-switch: subprocess_tree_kill_enabled=False -> the fallback path.
+# ---------------------------------------------------------------------------
+def test_tc07_tree_kill_off_passes_tree_kill_false(tmp_path, monkeypatch):
+    wt = tmp_path / "wt"
+    wt.mkdir()
+    monkeypatch.setattr(merge_gate, "get_worktree_path", lambda r, b: str(wt))
+    monkeypatch.setattr(merge_gate.settings, "subprocess_tree_kill_enabled", False, raising=False)
+    captured = {}
+
+    def _fake(cmd, *, cwd, timeout, env=None, grace_s=5.0, tree_kill=True):
+        captured["tree_kill"] = tree_kill
+        return ProcResult(returncode=0, stdout="1 passed", stderr="", timed_out=False)
+
+    monkeypatch.setattr(merge_gate, "run_in_process_group", _fake)
+    ok, reason = merge_gate.retest_branch(_REPO, _BRANCH)
+    assert ok is True and reason == "re-test green"
+    assert captured["tree_kill"] is False  # -> run_in_process_group degrades to subprocess.run
+
+
+# ---------------------------------------------------------------------------
+# D4 kill-switch: merge_retest_skip_when_current_enabled=False -> always re-test
+# after a rebase (even a proven no-op).
+# ---------------------------------------------------------------------------
+@pytest.fixture
+def gate_primitives(monkeypatch):
+    calls = {"retest": 0}
+    monkeypatch.setattr(qg.settings, "merge_gate_enabled", True, raising=False)
+    monkeypatch.setattr(qg.settings, "merge_gate_repos", "", raising=False)
+    monkeypatch.setattr(qg.settings, "premerge_rebase_always", True, raising=False)
+    monkeypatch.setattr(merge_gate, "acquire_merge_lease", lambda *a, **k: (True, "lease acquired"), raising=False)
+    monkeypatch.setattr(merge_gate, "branch_is_behind_main", lambda r, b: True, raising=False)
+    monkeypatch.setattr(merge_gate, "auto_rebase_onto_main", lambda r, b: (True, "rebased"), raising=False)
+    monkeypatch.setattr(merge_gate, "release_merge_lease", lambda *a, **k: None, raising=False)
+    # A PROVEN no-op rebase: HEAD is identical before/after.
+    monkeypatch.setattr(merge_gate, "head_sha", lambda r, b: "deadbeefcafe", raising=False)
+
+    def _retest(r, b):
+        calls["retest"] += 1
+        return True, "re-test green"
+
+    monkeypatch.setattr(merge_gate, "retest_branch", _retest, raising=False)
+    return calls
+
+
+def test_tc07_skip_when_current_off_always_retests(gate_primitives, monkeypatch):
+    monkeypatch.setattr(qg.settings, "merge_retest_skip_when_current_enabled", False, raising=False)
+    ok, reason = qg.check_branch_mergeable(_REPO, _WI, _BRANCH)
+    assert ok is True
+    assert reason == "rebased onto main, re-test green"
+    assert gate_primitives["retest"] == 1  # re-test STILL runs (D4 off)
+
+
+def test_tc07_skip_when_current_on_skips_noop_retest(gate_primitives, monkeypatch):
+    """Mirror sanity: with the flag ON, the proven no-op rebase skips the re-test."""
+    monkeypatch.setattr(qg.settings, "merge_retest_skip_when_current_enabled", True, raising=False)
+    ok, reason = qg.check_branch_mergeable(_REPO, _WI, _BRANCH)
+    assert ok is True
+    assert "re-test skipped" in reason
+    assert gate_primitives["retest"] == 0  # re-test NOT run on the no-op
+
+
+# ---------------------------------------------------------------------------
+# Non-self repo (enduro-trails): the merge-gate is N/A -> never the new paths.
+# ---------------------------------------------------------------------------
+def test_tc07_non_self_repo_is_noop(monkeypatch):
+    monkeypatch.setattr(qg.settings, "merge_gate_enabled", True, raising=False)
+    monkeypatch.setattr(qg.settings, "merge_gate_repos", "", raising=False)
+    # Guard: if the gate wrongly engaged it would touch the lease -> fail loudly.
+    monkeypatch.setattr(
+        merge_gate, "acquire_merge_lease",
+        lambda *a, **k: pytest.fail("merge-gate must be N/A for enduro-trails"),
+        raising=False,
+    )
+    ok, reason = qg.check_branch_mergeable("enduro-trails", "ET-1", "feature/ET-1-x")
+    assert ok is True
+    assert "N/A" in reason

tests/test_orch110_merge_gate_classify.py

@@ -0,0 +1,64 @@
+"""ORCH-110 TC-03: classify_retest_failure distinguishes infra-timeout from red.
+
+Covers D2 (FR-1 / AC-2): the pure predicate that lets the engine route an INFRA
+re-test timeout differently from a deterministically RED re-test, WITHOUT changing
+the name / PASS-FAIL semantics of the registered ``check_branch_mergeable``.
+
+The critical scope guard: an ``auto_rebase_onto_main`` "rebase timeout" is a
+DIFFERENT timeout (git hung) and must NOT be classified as the infra-tolerated
+re-test timeout (it stays on the rollback path).
+"""
+import os
+import tempfile
+
+os.environ.setdefault("ORCH_DB_PATH", os.path.join(tempfile.gettempdir(), "test_orch110_classify.db"))
+os.environ.setdefault("ORCH_GITEA_TOKEN", "test-token")
+os.environ.setdefault("ORCH_PLANE_API_TOKEN", "test-token")
+
+import pytest  # noqa: E402
+
+from src import merge_gate  # noqa: E402
+
+classify = merge_gate.classify_retest_failure
+
+
+@pytest.mark.parametrize(
+    "reason,expected",
+    [
+        ("re-test timeout after 900s", "timeout"),
+        ("re-test timeout after 600s", "timeout"),
+        ("re-test failed after rebase: 1 failed, 5 passed", "red"),
+        ("re-test failed: ...AssertionError\n1 failed", "red"),
+        ("merge-lock busy", "lock-busy"),
+        ("rebase conflict: src/db.py", "other"),
+        # SCOPE GUARD: a git "rebase timeout" is NOT the infra-tolerated re-test
+        # timeout — it must stay on the rollback path (ADR D2).
+        ("rebase timeout", "other"),
+        ("push --force-with-lease failed: ...", "other"),
+        ("", "other"),
+    ],
+)
+def test_tc03_classify_reasons(reason, expected):
+    assert classify(reason) == expected
+
+
+def test_tc03_classify_never_raises_on_bad_input():
+    # None / non-str must degrade to the safe "other" (-> rollback), never raise.
+    assert classify(None) == "other"
+    assert classify(12345) == "other"
+
+
+def test_tc03_case_insensitive():
+    assert classify("RE-TEST TIMEOUT AFTER 900S") == "timeout"
+    assert classify("Merge-Lock Busy") == "lock-busy"
+
+
+def test_tc03_distinct_from_lock_busy_and_conflict():
+    """timeout is a distinct class from the existing defer (lock-busy) and rollback
+    (conflict) reasons — the three must never collide."""
+    classes = {
+        classify("re-test timeout after 900s"),
+        classify("merge-lock busy"),
+        classify("rebase conflict: x"),
+    }
+    assert classes == {"timeout", "lock-busy", "other"}

tests/test_orch110_merge_gate_routing.py

@@ -0,0 +1,261 @@
+"""ORCH-110 TC-04 / TC-05 / TC-06 / TC-09: merge-gate infra-timeout routing.
+
+Drives the engine (``stage_engine.advance_stage``) on the deploy-staging -> deploy
+edge with ``check_branch_mergeable`` monkeypatched, exactly like the existing
+``test_stage_engine.TestMergeGate`` suite, and asserts the NEW routing (D3):
+
+  * TC-04 — an INFRA re-test timeout -> bounded infra-retry (re-queue the
+    staging-deployer with a delay, task STAYS on deploy-staging) — NOT a rollback to
+    development and NOT a developer-retry.
+  * TC-05 — a deterministically RED re-test STILL rolls back to development +
+    developer retry (BR-6 / AC-3 anti-over-tolerance).
+  * TC-06 — the infra-retry is bounded (anti-loop): after the budget it blocks with
+    ONE infra-alert, no infinite bounce, no new job, task NOT in development.
+  * TC-09 — never-raise: an error in the transient path is swallowed (a WARNING) and
+    never escapes into advance_stage.
+
+Offline: isolated sqlite DB; Plane/Telegram/notifications mocked at stage_engine.
+"""
+import os
+import tempfile
+
+import pytest
+
+_test_db = os.path.join(tempfile.gettempdir(), "test_orch110_routing.db")
+os.environ["ORCH_DB_PATH"] = _test_db
+os.environ["ORCH_REPOS_DIR"] = tempfile.gettempdir()
+os.environ.setdefault("ORCH_GITEA_TOKEN", "test-token")
+os.environ.setdefault("ORCH_PLANE_API_TOKEN", "test-token")
+
+from unittest.mock import MagicMock  # noqa: E402
+
+import src.db as _db  # noqa: E402
+from src.db import init_db, get_db  # noqa: E402
+from src import stage_engine  # noqa: E402
+from src.stage_engine import advance_stage  # noqa: E402
+
+_REPO = "orchestrator"
+_WI = "ORCH-110"
+_BRANCH = "feature/ORCH-110-x"
+_TIMEOUT_REASON = "re-test timeout after 900s"
+
+
+@pytest.fixture(autouse=True)
+def fresh_db(monkeypatch):
+    monkeypatch.setattr(_db.settings, "db_path", _test_db)
+    if os.path.exists(_test_db):
+        os.unlink(_test_db)
+    init_db()
+    yield
+
+
+@pytest.fixture(autouse=True)
+def silence_side_effects(monkeypatch):
+    for name in (
+        "notify_stage_change", "notify_qg_failure", "send_telegram",
+        "plane_notify_stage", "plane_notify_qg", "plane_add_comment",
+        "set_issue_in_progress", "set_issue_blocked",
+    ):
+        monkeypatch.setattr(stage_engine, name, MagicMock())
+    # The merge-gate sub-gate runs only AFTER the stage QG + security + coverage pass;
+    # the self-deploy Phase A interception is irrelevant (merge-gate intervenes first).
+    monkeypatch.setattr(stage_engine.settings, "deploy_require_manual_approve", False)
+
+
+@pytest.fixture(autouse=True)
+def tolerance_on(monkeypatch):
+    monkeypatch.setattr(stage_engine.settings, "merge_retest_infra_tolerance_enabled", True)
+    monkeypatch.setattr(stage_engine.settings, "merge_retest_infra_max_retries", 2)
+    monkeypatch.setattr(stage_engine.settings, "merge_retest_infra_retry_delay_s", 120)
+
+
+def _pass(*a, **k):
+    return (True, "ok")
+
+
+def _fail(reason):
+    def _f(*a, **k):
+        return (False, reason)
+    return _f
+
+
+def _patch_gates(monkeypatch, merge_reason):
+    monkeypatch.setattr(
+        stage_engine, "QG_CHECKS",
+        {**stage_engine.QG_CHECKS,
+         "check_staging_status": _pass,
+         "check_security_gate": _pass,
+         "check_coverage_gate": _pass,
+         "check_branch_mergeable": _fail(merge_reason),
+         "check_staging_image_fresh": _pass},
+    )
+
+
+def _make_task(stage="deploy-staging"):
+    conn = get_db()
+    cur = conn.execute(
+        "INSERT INTO tasks (plane_id, work_item_id, repo, branch, stage) VALUES (?,?,?,?,?)",
+        (f"plane-{_WI}", _WI, _REPO, _BRANCH, stage),
+    )
+    tid = cur.lastrowid
+    conn.commit()
+    conn.close()
+    return tid
+
+
+def _stage(task_id):
+    conn = get_db()
+    row = conn.execute("SELECT stage FROM tasks WHERE id=?", (task_id,)).fetchone()
+    conn.close()
+    return row[0]
+
+
+def _jobs():
+    conn = get_db()
+    rows = conn.execute(
+        "SELECT agent, task_content, available_at FROM jobs ORDER BY id"
+    ).fetchall()
+    conn.close()
+    return [dict(r) for r in rows]
+
+
+def _seed_infra_retry_jobs(task_id, n):
+    conn = get_db()
+    for _ in range(n):
+        conn.execute(
+            "INSERT INTO jobs (agent, repo, task_id, task_content) "
+            "VALUES ('deployer','orchestrator',?, 'Note: merge-gate infra-timeout retry')",
+            (task_id,),
+        )
+    conn.commit()
+    conn.close()
+
+
+def _advance(task_id):
+    return advance_stage(
+        task_id, "deploy-staging", _REPO, _WI, _BRANCH, finished_agent="deployer"
+    )
+
+
+# ---------------------------------------------------------------------------
+# TC-04 — infra-timeout -> bounded infra-retry, NOT a rollback to development.
+# ---------------------------------------------------------------------------
+def test_tc04_infra_timeout_reschedules_not_rollback(monkeypatch):
+    _patch_gates(monkeypatch, _TIMEOUT_REASON)
+    task_id = _make_task()
+    res = _advance(task_id)
+
+    assert res.advanced is False
+    assert res.rolled_back_to is None             # NOT a code-fault rollback
+    assert res.note == "merge-gate-infra-retry"
+    assert _stage(task_id) == "deploy-staging"    # stays put
+    jobs = _jobs()
+    assert len(jobs) == 1
+    assert jobs[0]["agent"] == "deployer"          # re-queued staging-deployer, NOT developer
+    assert "merge-gate infra-timeout retry" in jobs[0]["task_content"]
+    assert jobs[0]["available_at"] is not None     # delayed re-pickup
+    assert stage_engine.set_issue_blocked.called is False
+    # No developer-retry semantics: the rollback comment / in-progress is never set.
+    assert stage_engine.set_issue_in_progress.called is False
+
+
+def test_tc04_killswitch_within_routing_is_observed(monkeypatch):
+    """The infra-timeout always bumps the timeout counter (observability), even when
+    routed to the retry path."""
+    _patch_gates(monkeypatch, _TIMEOUT_REASON)
+    before = stage_engine.merge_gate.merge_gate_status()["retest_timeout_total"]
+    task_id = _make_task()
+    _advance(task_id)
+    after = stage_engine.merge_gate.merge_gate_status()["retest_timeout_total"]
+    assert after == before + 1
+
+
+# ---------------------------------------------------------------------------
+# TC-05 — a deterministically RED re-test STILL rolls back (BR-6 / AC-3).
+# ---------------------------------------------------------------------------
+def test_tc05_red_retest_still_rolls_back(monkeypatch):
+    _patch_gates(monkeypatch, "re-test failed after rebase: 1 failed, 5 passed")
+    task_id = _make_task()
+    res = _advance(task_id)
+
+    assert res.advanced is False
+    assert res.rolled_back_to == "development"
+    assert _stage(task_id) == "development"
+    jobs = _jobs()
+    assert len(jobs) == 1
+    assert jobs[0]["agent"] == "developer"        # developer re-queued (retry)
+
+
+def test_tc05_conflict_still_rolls_back(monkeypatch):
+    _patch_gates(monkeypatch, "rebase conflict: src/db.py")
+    task_id = _make_task()
+    res = _advance(task_id)
+    assert res.rolled_back_to == "development"
+    assert _stage(task_id) == "development"
+    assert _jobs()[0]["agent"] == "developer"
+
+
+# ---------------------------------------------------------------------------
+# TC-06 — anti-loop: infra-retry is bounded; exhaustion -> ONE infra-alert.
+# ---------------------------------------------------------------------------
+def test_tc06_infra_retry_bounded_then_infra_alert(monkeypatch):
+    monkeypatch.setattr(stage_engine.settings, "merge_retest_infra_max_retries", 2)
+    _patch_gates(monkeypatch, _TIMEOUT_REASON)
+    task_id = _make_task()
+    _seed_infra_retry_jobs(task_id, 2)  # budget already spent
+
+    res = _advance(task_id)
+    assert res.advanced is False
+    assert res.rolled_back_to is None                  # NOT a rollback even at exhaustion
+    assert res.note == "merge-gate-infra-retry-exhausted"
+    assert res.alerted is True
+    assert _stage(task_id) == "deploy-staging"         # NOT moved to development
+    assert stage_engine.set_issue_blocked.called
+    assert stage_engine.send_telegram.called
+    # No NEW retry job past the cap (still only the 2 we seeded).
+    assert len(_jobs()) == 2
+    # The alert is INFRA-specific, not "developer must fix".
+    msg = stage_engine.send_telegram.call_args[0][0]
+    assert "infra" in msg.lower() or "ресурс" in msg.lower()
+    assert "НЕ дефект кода" in msg
+
+
+def test_tc06_below_budget_keeps_retrying(monkeypatch):
+    monkeypatch.setattr(stage_engine.settings, "merge_retest_infra_max_retries", 3)
+    _patch_gates(monkeypatch, _TIMEOUT_REASON)
+    task_id = _make_task()
+    _seed_infra_retry_jobs(task_id, 1)  # one retry already done, budget 3
+
+    res = _advance(task_id)
+    assert res.note == "merge-gate-infra-retry"
+    assert res.alerted is not True
+    # The seeded job + the new retry job.
+    assert len(_jobs()) == 2
+
+
+# ---------------------------------------------------------------------------
+# TC-09 — never-raise: an error in the transient path is swallowed.
+# ---------------------------------------------------------------------------
+def test_tc09_infra_retry_never_raises(monkeypatch):
+    _patch_gates(monkeypatch, _TIMEOUT_REASON)
+
+    def _boom(*a, **k):
+        raise RuntimeError("enqueue exploded")
+
+    monkeypatch.setattr(stage_engine, "enqueue_job", _boom)
+    task_id = _make_task()
+    # Must NOT raise into advance_stage.
+    res = _advance(task_id)
+    assert res.note == "merge-gate-infra-retry-error"
+    assert _stage(task_id) == "deploy-staging"  # left for the reconciler/reaper
+
+
+def test_tc09_killswitch_off_falls_back_to_rollback(monkeypatch):
+    """tolerance off -> a timeout takes the prior rollback path byte-for-byte (NFR-2)."""
+    monkeypatch.setattr(stage_engine.settings, "merge_retest_infra_tolerance_enabled", False)
+    _patch_gates(monkeypatch, _TIMEOUT_REASON)
+    task_id = _make_task()
+    res = _advance(task_id)
+    assert res.rolled_back_to == "development"
+    assert _stage(task_id) == "development"
+    assert _jobs()[0]["agent"] == "developer"

tests/test_orch110_observability.py

@@ -0,0 +1,135 @@
+"""ORCH-110 TC-12: observability of the infra-timeout path (D6 / FR-6 / AC-9).
+
+The infra-timeout state must be (a) reflected in read-only counters, (b) surfaced in
+the additive ``merge_gate`` block of ``GET /queue``, and (c) distinguishable from a
+code-fault rollback — with the exhaustion alert carrying the CLICKABLE issue number
+and an explicitly infrastructural (NOT "developer must fix") wording. No dedup/overlap
+with ORCH-111 (which only OBSERVES surviving processes; ORCH-110 prevents/tolerates).
+"""
+import asyncio
+import os
+import tempfile
+
+_test_db = os.path.join(tempfile.gettempdir(), "test_orch110_observability.db")
+os.environ["ORCH_DB_PATH"] = _test_db
+os.environ["ORCH_REPOS_DIR"] = tempfile.gettempdir()
+os.environ.setdefault("ORCH_GITEA_TOKEN", "test-token")
+os.environ.setdefault("ORCH_PLANE_API_TOKEN", "test-token")
+
+from unittest.mock import MagicMock  # noqa: E402
+
+import pytest  # noqa: E402
+
+import src.db as _db  # noqa: E402
+from src.db import init_db, get_db  # noqa: E402
+from src import merge_gate  # noqa: E402
+from src import stage_engine  # noqa: E402
+from src.stage_engine import AdvanceResult  # noqa: E402
+
+
+@pytest.fixture(autouse=True)
+def fresh_db(monkeypatch):
+    monkeypatch.setattr(_db.settings, "db_path", _test_db)
+    if os.path.exists(_test_db):
+        os.unlink(_test_db)
+    init_db()
+    # Reset the in-process counters so each test starts from a known baseline.
+    merge_gate._MERGE_GATE_COUNTERS.update(
+        retest_timeout_total=0, retest_infra_retry_total=0,
+        retest_infra_exhausted_total=0, retest_skipped_current_total=0,
+        last_infra_timeout_wi=None,
+    )
+    yield
+
+
+# ---------------------------------------------------------------------------
+# merge_gate_status() snapshot + counter increments.
+# ---------------------------------------------------------------------------
+def test_tc12_status_exposes_flags_and_counters():
+    snap = merge_gate.merge_gate_status()
+    for key in (
+        "infra_tolerance_enabled", "infra_max_retries", "infra_retry_delay_s",
+        "skip_when_current_enabled", "tree_kill_enabled", "retest_timeout_s",
+        "retest_timeout_total", "retest_infra_retry_total",
+        "retest_infra_exhausted_total", "retest_skipped_current_total",
+        "last_infra_timeout_wi",
+    ):
+        assert key in snap, f"missing /queue merge_gate key: {key}"
+
+
+def test_tc12_counters_track_infra_timeout_distinctly():
+    merge_gate.note_retest_timeout("ORCH-110")
+    merge_gate.note_retest_infra_retry()
+    merge_gate.note_retest_infra_exhausted()
+    merge_gate.note_retest_skipped_current()
+    snap = merge_gate.merge_gate_status()
+    assert snap["retest_timeout_total"] == 1
+    assert snap["retest_infra_retry_total"] == 1
+    assert snap["retest_infra_exhausted_total"] == 1
+    assert snap["retest_skipped_current_total"] == 1
+    # Distinguishable from a code-fault: the last infra-timeout WI is tracked here,
+    # NOT in the merge-verify (code/merge) counters.
+    assert snap["last_infra_timeout_wi"] == "ORCH-110"
+
+
+def test_tc12_status_never_raises(monkeypatch):
+    # A broken settings attribute -> the snapshot degrades, never raises.
+    monkeypatch.delattr(merge_gate.settings, "merge_retest_infra_tolerance_enabled", raising=False)
+    snap = merge_gate.merge_gate_status()
+    assert isinstance(snap, dict)
+
+
+# ---------------------------------------------------------------------------
+# GET /queue carries the additive merge_gate block (and the legacy keys stay).
+# ---------------------------------------------------------------------------
+def test_tc12_queue_endpoint_includes_merge_gate_block():
+    from src.main import queue
+    out = asyncio.run(queue())
+    assert "merge_gate" in out
+    assert "infra_tolerance_enabled" in out["merge_gate"]
+    # The pre-existing observability keys are untouched (additive only).
+    assert "merge_verify" in out and "coverage" in out
+
+
+# ---------------------------------------------------------------------------
+# The exhaustion alert is INFRA-specific + carries the clickable issue number,
+# distinct from the code-fault "Merge-gate still failing after N developer retries".
+# ---------------------------------------------------------------------------
+def test_tc12_exhaustion_alert_is_infra_specific_and_clickable(monkeypatch):
+    sent = {}
+
+    def _tg(msg):
+        sent["msg"] = msg
+
+    monkeypatch.setattr(stage_engine, "send_telegram", _tg)
+    monkeypatch.setattr(stage_engine, "set_issue_blocked", MagicMock())
+    monkeypatch.setattr(stage_engine, "plane_add_comment", MagicMock())
+    # link_for builds the clickable number; use a recognisable sentinel.
+    monkeypatch.setattr(stage_engine, "link_for", lambda wi, **k: f"<a href='x'>{wi}</a>")
+    monkeypatch.setattr(stage_engine.settings, "merge_retest_infra_max_retries", 2)
+
+    # Seed an exhausted budget for the task.
+    conn = get_db()
+    cur = conn.execute(
+        "INSERT INTO tasks (plane_id, work_item_id, repo, branch, stage) VALUES (?,?,?,?,?)",
+        ("plane-ORCH-110", "ORCH-110", "orchestrator", "feature/ORCH-110-x", "deploy-staging"),
+    )
+    task_id = cur.lastrowid
+    for _ in range(2):
+        conn.execute(
+            "INSERT INTO jobs (agent, repo, task_id, task_content) "
+            "VALUES ('deployer','orchestrator',?, 'Note: merge-gate infra-timeout retry')",
+            (task_id,),
+        )
+    conn.commit()
+    conn.close()
+
+    res = AdvanceResult()
+    stage_engine._handle_merge_gate_infra_retry(
+        task_id, "deploy-staging", "orchestrator", "ORCH-110",
+        "feature/ORCH-110-x", "re-test timeout after 900s", res,
+    )
+    assert res.note == "merge-gate-infra-retry-exhausted"
+    assert "ORCH-110" in sent["msg"]                  # clickable issue number present
+    assert "developer retries" not in sent["msg"]     # NOT the code-fault wording
+    assert "НЕ дефект кода" in sent["msg"]             # explicitly infrastructural

tests/test_orch110_retest_contract.py

@@ -0,0 +1,95 @@
+"""ORCH-110 TC-11: the local re-test necessity contract (D4 / FR-4 / AC-6).
+
+The chosen contract: run the local re-test IFF the pre-merge rebase actually moved
+HEAD (``main`` had moved -> a real semantic-conflict risk). When the rebase is a
+PROVEN no-op (HEAD unchanged -> branch already at origin/main, the CI/tester/staging-
+validated commit) the re-test is SKIPPED — it would be a redundant single point of
+false failure. On ANY uncertainty (an empty SHA) the re-test runs (fail-safe -> the
+red-rollback contract BR-6/AC-3 is never weakened).
+"""
+import os
+import tempfile
+
+os.environ.setdefault("ORCH_DB_PATH", os.path.join(tempfile.gettempdir(), "test_orch110_contract.db"))
+os.environ.setdefault("ORCH_REPOS_DIR", tempfile.gettempdir())
+os.environ.setdefault("ORCH_GITEA_TOKEN", "test-token")
+os.environ.setdefault("ORCH_PLANE_API_TOKEN", "test-token")
+
+import pytest  # noqa: E402
+
+from src import merge_gate  # noqa: E402
+from src.qg import checks as qg  # noqa: E402
+
+_REPO = "orchestrator"
+_BRANCH = "feature/ORCH-110-x"
+_WI = "ORCH-110"
+
+
+@pytest.fixture
+def gate(monkeypatch):
+    """Real check_branch_mergeable, mocked primitives; record control flow."""
+    state = {"retest": 0, "released": 0, "rebased": 0}
+    monkeypatch.setattr(qg.settings, "merge_gate_enabled", True, raising=False)
+    monkeypatch.setattr(qg.settings, "merge_gate_repos", "", raising=False)
+    monkeypatch.setattr(qg.settings, "premerge_rebase_always", True, raising=False)
+    monkeypatch.setattr(qg.settings, "merge_retest_skip_when_current_enabled", True, raising=False)
+    monkeypatch.setattr(merge_gate, "acquire_merge_lease", lambda *a, **k: (True, "lease acquired"))
+    monkeypatch.setattr(merge_gate, "branch_is_behind_main", lambda r, b: True)
+
+    def _rebase(r, b):
+        state["rebased"] += 1
+        return True, "rebased onto origin/main"
+
+    def _release(r, b=None):
+        state["released"] += 1
+
+    def _retest(r, b):
+        state["retest"] += 1
+        return True, "re-test green"
+
+    monkeypatch.setattr(merge_gate, "auto_rebase_onto_main", _rebase)
+    monkeypatch.setattr(merge_gate, "release_merge_lease", _release)
+    monkeypatch.setattr(merge_gate, "retest_branch", _retest)
+    return state
+
+
+def _set_head_shas(monkeypatch, shas):
+    seq = list(shas)
+    monkeypatch.setattr(merge_gate, "head_sha", lambda r, b: seq.pop(0) if seq else "")
+
+
+def test_tc11_noop_rebase_skips_retest_lease_held(gate, monkeypatch):
+    """Proven no-op rebase (HEAD unchanged) -> skip re-test, PASS, lease HELD."""
+    _set_head_shas(monkeypatch, ["sha_same", "sha_same"])
+    ok, reason = qg.check_branch_mergeable(_REPO, _WI, _BRANCH)
+    assert ok is True
+    assert "re-test skipped" in reason
+    assert gate["retest"] == 0          # re-test NOT run
+    assert gate["released"] == 0        # lease HELD until the merge
+
+
+def test_tc11_head_moved_runs_retest(gate, monkeypatch):
+    """A real catch-up (HEAD moved) -> re-test RUNS (the ORCH-043 risk is real)."""
+    _set_head_shas(monkeypatch, ["sha_old", "sha_new"])
+    ok, reason = qg.check_branch_mergeable(_REPO, _WI, _BRANCH)
+    assert ok is True
+    assert reason == "rebased onto main, re-test green"
+    assert gate["retest"] == 1
+
+
+@pytest.mark.parametrize("shas", [["", "sha_new"], ["sha_old", ""], ["", ""]])
+def test_tc11_uncertain_sha_runs_retest_failsafe(gate, monkeypatch, shas):
+    """Cannot prove a no-op (empty SHA) -> re-test RUNS (fail-safe, never skips on
+    uncertainty)."""
+    _set_head_shas(monkeypatch, shas)
+    ok, reason = qg.check_branch_mergeable(_REPO, _WI, _BRANCH)
+    assert ok is True
+    assert gate["retest"] == 1          # re-test still runs on uncertainty
+
+
+def test_tc11_skip_bumps_observability_counter(gate, monkeypatch):
+    """The skip increments the read-only observability counter (D6)."""
+    merge_gate._MERGE_GATE_COUNTERS["retest_skipped_current_total"] = 0
+    _set_head_shas(monkeypatch, ["sha_same", "sha_same"])
+    qg.check_branch_mergeable(_REPO, _WI, _BRANCH)
+    assert merge_gate._MERGE_GATE_COUNTERS["retest_skipped_current_total"] == 1

tests/test_orch110_retest_lifecycle.py

@@ -0,0 +1,185 @@
+"""ORCH-110 TC-01 / TC-02: process-group tree-kill of orchestrator-spawned pytest.
+
+Covers D1 (FR-2 / BR-3 / AC-4): a timeout in ``merge_gate.retest_branch`` /
+``coverage_gate.measure_coverage`` must tree-kill the WHOLE subprocess subtree
+(children + grandchildren) — the orphan-leak root cause of the ORCH-109 incident —
+not just the direct child.
+
+Fully deterministic and offline:
+  * TC-01 runs a REAL throwaway process that spawns a long-sleeping grandchild via
+    ``sys.executable`` (always present), then asserts the grandchild is dead after a
+    ``run_in_process_group`` timeout — the concrete "no orphan survives" proof.
+  * TC-01b / TC-02 assert the two call sites delegate to ``run_in_process_group``
+    with the right tree-kill kwargs and map a timed-out ProcResult to their existing
+    contract (``(False, "re-test timeout after <T>s")`` / ``None``).
+No network, no Plane/Gitea, no Claude CLI.
+"""
+import os
+import tempfile
+import time
+
+os.environ.setdefault("ORCH_DB_PATH", os.path.join(tempfile.gettempdir(), "test_orch110_lifecycle.db"))
+os.environ.setdefault("ORCH_REPOS_DIR", tempfile.gettempdir())
+os.environ.setdefault("ORCH_GITEA_TOKEN", "test-token")
+os.environ.setdefault("ORCH_PLANE_API_TOKEN", "test-token")
+
+import sys  # noqa: E402
+
+import pytest  # noqa: E402
+
+from src import coverage_gate as cg  # noqa: E402
+from src import merge_gate  # noqa: E402
+from src import proc_group  # noqa: E402
+from src.proc_group import ProcResult, run_in_process_group  # noqa: E402
+
+_POSIX = hasattr(os, "killpg") and hasattr(os, "setsid")
+
+
+def _alive(pid: int) -> bool:
+    """True iff ``pid`` is a live (non-reaped) process."""
+    try:
+        os.kill(pid, 0)
+        return True
+    except ProcessLookupError:
+        return False
+    except PermissionError:
+        return True
+
+
+# ---------------------------------------------------------------------------
+# TC-01: real grandchild is tree-killed on timeout (the FR-2 / AC-4 proof).
+# ---------------------------------------------------------------------------
+@pytest.mark.skipif(not _POSIX, reason="process-group tree-kill is POSIX-only")
+def test_tc01_grandchild_orphan_is_tree_killed_on_timeout(tmp_path):
+    pidfile = tmp_path / "grandchild.pid"
+    script = tmp_path / "spawner.py"
+    # The leader spawns a long-sleeping grandchild, records its pid, then sleeps far
+    # past the timeout. A naive subprocess.run(timeout=) would kill ONLY the leader
+    # and leak the grandchild (the ORCH-109 bug). The process-group tree-kill must
+    # reap the grandchild too.
+    script.write_text(
+        "import subprocess, sys, time\n"
+        "gc = subprocess.Popen([sys.executable, '-c', 'import time; time.sleep(120)'])\n"
+        "open(sys.argv[1], 'w').write(str(gc.pid))\n"
+        "time.sleep(120)\n",
+        encoding="utf-8",
+    )
+
+    res = run_in_process_group(
+        [sys.executable, str(script), str(pidfile)],
+        cwd=str(tmp_path), timeout=2, grace_s=1, tree_kill=True,
+    )
+    assert res.timed_out is True
+
+    # Wait for the leader to actually have written the grandchild pid.
+    deadline = time.time() + 5
+    while time.time() < deadline and not pidfile.exists():
+        time.sleep(0.05)
+    assert pidfile.exists(), "spawner never recorded the grandchild pid"
+    gc_pid = int(pidfile.read_text().strip())
+
+    # The grandchild must die (be reaped) within a short window after the tree-kill.
+    deadline = time.time() + 8
+    while time.time() < deadline and _alive(gc_pid):
+        time.sleep(0.1)
+    assert not _alive(gc_pid), "grandchild orphan survived the timeout tree-kill (FR-2/AC-4)"
+
+
+@pytest.mark.skipif(not _POSIX, reason="process-group runner is POSIX-only")
+def test_tc01_green_and_red_returncodes_passthrough(tmp_path):
+    """The runner returns the real exit code on a normal (non-timeout) completion."""
+    ok = run_in_process_group(
+        [sys.executable, "-c", "import sys; sys.exit(0)"],
+        cwd=str(tmp_path), timeout=30, tree_kill=True,
+    )
+    assert ok.timed_out is False and ok.returncode == 0
+    red = run_in_process_group(
+        [sys.executable, "-c", "import sys; sys.exit(3)"],
+        cwd=str(tmp_path), timeout=30, tree_kill=True,
+    )
+    assert red.timed_out is False and red.returncode == 3
+
+
+def test_tc01_fallback_when_tree_kill_disabled(tmp_path, monkeypatch):
+    """tree_kill=False -> degrades to subprocess.run (never-break); still times out."""
+    res = run_in_process_group(
+        [sys.executable, "-c", "import time; time.sleep(30)"],
+        cwd=str(tmp_path), timeout=1, tree_kill=False,
+    )
+    assert res.timed_out is True
+
+
+def test_tc01_fallback_on_non_posix(tmp_path, monkeypatch):
+    """No os.killpg/setsid (non-POSIX) -> the plain subprocess.run fallback runs."""
+    monkeypatch.setattr(proc_group, "_tree_kill_supported", lambda: False)
+    res = run_in_process_group(
+        [sys.executable, "-c", "import sys; sys.exit(0)"],
+        cwd=str(tmp_path), timeout=30,
+    )
+    assert res.timed_out is False and res.returncode == 0
+
+
+# ---------------------------------------------------------------------------
+# TC-01b: retest_branch delegates to the process-group runner with the right
+# kwargs and maps a timeout to its existing contract.
+# ---------------------------------------------------------------------------
+def test_tc01b_retest_branch_uses_process_group_and_maps_timeout(tmp_path, monkeypatch):
+    wt = tmp_path / "wt"
+    wt.mkdir()
+    monkeypatch.setattr(merge_gate, "get_worktree_path", lambda repo, branch: str(wt))
+    monkeypatch.setattr(merge_gate.settings, "merge_retest_timeout_s", 900, raising=False)
+    monkeypatch.setattr(merge_gate.settings, "subprocess_tree_kill_enabled", True, raising=False)
+    monkeypatch.setattr(merge_gate.settings, "agent_kill_grace_seconds", 7, raising=False)
+
+    captured = {}
+
+    def _fake(cmd, *, cwd, timeout, env=None, grace_s=5.0, tree_kill=True):
+        captured.update(cmd=cmd, cwd=cwd, timeout=timeout, grace_s=grace_s, tree_kill=tree_kill)
+        return ProcResult(returncode=None, stdout="", stderr="", timed_out=True)
+
+    monkeypatch.setattr(merge_gate, "run_in_process_group", _fake)
+    ok, reason = merge_gate.retest_branch("orchestrator", "feature/x")
+    assert ok is False
+    assert reason == "re-test timeout after 900s"
+    # FR-2: spawned via the process-group runner with tree-kill + the agent grace.
+    assert captured["tree_kill"] is True
+    assert captured["grace_s"] == 7
+    assert captured["cmd"][:3] == ["python", "-m", "pytest"]
+
+
+# ---------------------------------------------------------------------------
+# TC-02: coverage_gate.measure_coverage — sibling orphan source (BR-3). Timeout
+# tree-kills + returns None; uses the same runner with tree-kill from settings.
+# ---------------------------------------------------------------------------
+def test_tc02_measure_coverage_timeout_tree_kill_returns_none(tmp_path, monkeypatch):
+    monkeypatch.setattr(cg, "ensure_worktree", lambda r, b: str(tmp_path))
+    monkeypatch.setattr(cg.settings, "subprocess_tree_kill_enabled", True, raising=False)
+    monkeypatch.setattr(cg.settings, "agent_kill_grace_seconds", 5, raising=False)
+    monkeypatch.setattr(cg.settings, "coverage_run_timeout_s", 900, raising=False)
+
+    captured = {}
+
+    def _fake(cmd, *, cwd, timeout, env=None, grace_s=5.0, tree_kill=True):
+        captured.update(tree_kill=tree_kill, grace_s=grace_s, timeout=timeout)
+        return ProcResult(returncode=None, stdout="", stderr="", timed_out=True)
+
+    monkeypatch.setattr(cg, "run_in_process_group", _fake)
+    assert cg.measure_coverage("orchestrator", "feature/x") is None
+    assert captured["tree_kill"] is True
+    assert captured["grace_s"] == 5
+    assert captured["timeout"] == 900
+
+
+def test_tc02_measure_coverage_respects_tree_kill_killswitch(tmp_path, monkeypatch):
+    """With the tree-kill kill-switch off, the runner is asked for the fallback path."""
+    monkeypatch.setattr(cg, "ensure_worktree", lambda r, b: str(tmp_path))
+    monkeypatch.setattr(cg.settings, "subprocess_tree_kill_enabled", False, raising=False)
+    captured = {}
+
+    def _fake(cmd, *, cwd, timeout, env=None, grace_s=5.0, tree_kill=True):
+        captured.update(tree_kill=tree_kill)
+        return ProcResult(returncode=None, stdout="", stderr="", timed_out=True)
+
+    monkeypatch.setattr(cg, "run_in_process_group", _fake)
+    assert cg.measure_coverage("orchestrator", "feature/x") is None
+    assert captured["tree_kill"] is False