fix(coordination): make Stage 2C capture and rollback recoverable

.github/workflows/python-tests.yml

@@ -117,7 +117,7 @@ jobs:
         # Without timing history least_duration alternates equal-weight tests.
         # Each runner retains the measured two-worker pool.
         run: >-
-          python -m pytest -q -n 2
+          python -m pytest -q -n 2 -m "not stage2c_e2e"
           --splits 2 --group ${{ matrix.shard }}
           --splitting-algorithm least_duration
           --durations=25 --durations-min=1
@@ -176,6 +176,70 @@ jobs:
     secrets:
       SONAR_TOKEN: ${{ secrets.SONAR_TOKEN }}
 
+  stage2c-correctness-e2e:
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    steps:
+      - uses: actions/checkout@v7
+      - uses: actions/setup-python@v6
+        with:
+          python-version: "3.11"
+          cache: pip
+          cache-dependency-path: tests/requirements-stage2c-linux-py311.txt
+      - uses: actions/setup-node@v6
+        with:
+          node-version: "22.6"
+          cache: npm
+      - name: Install locked test and package build tools
+        run: |
+          python -m pip install --disable-pip-version-check --require-hashes --only-binary=:all: -r tests/requirements-stage2c-linux-py311.txt
+          npm ci --ignore-scripts
+      - name: Build and verify the checked-out source package
+        run: |
+          python -m build --no-isolation --wheel --outdir .local/stage2c-source-install
+          python - <<'PYTHON'
+          import hashlib
+          from pathlib import Path
+
+          directory = Path(".local/stage2c-source-install").resolve()
+          wheels = list(directory.glob("*.whl"))
+          if len(wheels) != 1:
+              raise SystemExit("Expected exactly one wheel from the checked-out source")
+          wheel = wheels[0]
+          digest = hashlib.sha256(wheel.read_bytes()).hexdigest()
+          (directory / "requirements.txt").write_text(
+              f"loopx @ {wheel.as_uri()} --hash=sha256:{digest}\n", encoding="utf-8"
+          )
+          PYTHON
+          python -m pip install --disable-pip-version-check --require-hashes --no-deps --no-index -r .local/stage2c-source-install/requirements.txt
+          python -m pip check
+          # Remove the generated source copy before pytest's normal discovery.
+          python -c "import shutil; shutil.rmtree('build')"
+      - name: Qualify real CLI, mixed writers, process death, and recovery
+        env:
+          LOOPX_SHADOW_COMPARISON_OUTPUT: .local/stage2c-observables
+        run: python -m pytest -q -m stage2c_e2e --junitxml=stage2c-e2e.xml
+      - name: Reject deliberate correctness regressions
+        run: python examples/shared-goal-authority-e2e/mutants.py --output .local/stage2c-mutants
+      - name: Build independently installed distributions
+        run: python -m build --no-isolation
+      - name: Qualify wheel outside the repository
+        run: python examples/shared-goal-authority-e2e/installed.py --artifact dist/*.whl --report-json installed-wheel.json
+      - name: Qualify sdist outside the repository
+        run: python examples/shared-goal-authority-e2e/installed.py --artifact dist/*.tar.gz --report-json installed-sdist.json
+      - name: Retain bounded acceptance evidence
+        if: always()
+        uses: actions/upload-artifact@v7
+        with:
+          name: stage2c-correctness-evidence
+          include-hidden-files: true
+          path: |
+            stage2c-e2e.xml
+            .local/stage2c-observables/
+            installed-wheel.json
+            installed-sdist.json
+            .local/stage2c-mutants/
+
   windows-powershell:
     runs-on: windows-latest
     timeout-minutes: 20

docs/architecture/rfcs/shared-goal-authority-state-provider-v0.md

@@ -2268,9 +2268,14 @@ the unresolved gates remain prerequisites for a real promotion.
   writer fence bound to that revision; provider-first `mutate` and
   `todo_read` that never fall back to Markdown.
 - The fence integration: every Python Todo mutation and every native task-lease
-  acquire, renew, transfer, and release checks the durable fence while holding
-  its own lock; an absent fence costs no runtime call; a present, unreadable,
-  or invalid fence fails closed.
+  acquire, renew, transfer, release, verify, and committed releasing
+  fence-close checks the durable fence while holding its own lock; an absent
+  fence costs no runtime call; a present, unreadable, or invalid fence fails
+  closed. A fenced write is rejected before its first side effect as a
+  validation-stage `permission_denied` with no receipt; the shared check owns
+  only the typed reason and the fence binding facts, and each caller adapter
+  renders one provider-neutral remediation from them and carries the check
+  result under `write_check`.
 - The complete Todo read model: `loopx_todo_canonical_read_record_v0` publishes
   a versioned field manifest, and the TypeScript projection rejects a
   replacement that drops fields already present on a stored record.

docs/architecture/rfcs/shared-goal-authority-state-provider-v0.zh-CN.md

@@ -1799,8 +1799,11 @@ authority 语义与各 provider 的物理 retention 策略。本文档不实现
   writer fence 的 `coordination.local_authority.promote`；永不回退到 Markdown 的
   provider-first `mutate` 与 `todo_read`。
 - fence 集成：每个 Python Todo mutation 与每个 native task-lease
-  acquire/renew/transfer/release 都在自己的锁内检查持久 fence；fence 不存在时零运行时
-  调用；fence 存在但不可读或无效时 fail closed。
+  acquire/renew/transfer/release/verify 以及已提交的释放型 fence-close 都在自己的锁内
+  检查持久 fence；fence 不存在时零运行时调用；fence 存在但不可读或无效时 fail closed。
+  被 fence 拒绝的写在第一个副作用之前以 validation 阶段的 `permission_denied`
+  返回且不产生回执；共享检查只拥有类型化原因与 fence 绑定事实，各调用方 adapter
+  据此渲染同一条 provider 中立的 remediation，并把检查结果放在 `write_check` 下携带。
 - 完整 Todo read model：`loopx_todo_canonical_read_record_v0` 发布带版本字段 manifest；
   TypeScript projection 拒绝 replacement 丢弃既有记录中已经存在的字段。
 

examples/shared-goal-authority-e2e/README.md

@@ -84,23 +84,33 @@ confined to the `bindings` block nulls every binding, marks
 `summary.privacy_violations`, which no flag relaxes. Evidence therefore
 carries counters, cursors, outcome tokens, and sha256 prefixes only.
 
-## Test seams later PRs must provide
-
-The pending `s2c2.*` rows will be implemented against these seams; a Stage 2C
-parity PR that does not expose them cannot be ladder-verified:
-
-- a drain lock file at `<runtime>/authority-shadow/outbox/<goal>/drain` so the
-  ladder can hold the drain window with `loopx.file_lock.exclusive_file_lock`
-  exactly as it holds `<runtime>/authority-shadow/file/<goal>/observation`
-  today, then SIGKILL a writer before or during drain;
-- one file per outbox entry under `<runtime>/authority-shadow/outbox/<goal>/`
-  with a prepared-then-committed marker, so pending entries are countable and
-  a rollback with pending entries is observable from disk;
-- `drain` and `verify` product commands that emit JSON with `drained_count`,
-  `cursor_before`, `cursor_after`, `parity_verdict`, and the source and
-  candidate digests, so parity-equal and foreign-edit rows can assert on
-  typed fields rather than prose;
-- the same commands must resolve the runtime root the way `todo` and
-  `task-lease` do (`effective_runtime_root`), so the one-lineage guarantee that
-  `s2c1.dual_runtime_root_consistency` proves for the observation hooks also
-  holds for drain and verify.
+## Bounded outbox correctness and the pending ladder
+
+The independently runnable [correctness suite](correctness.md) covers the
+`file_outbox_v1` capture lineage, strict cursor recovery, mixed writers,
+source fences, and recoverable bootstrap/rollback. It uses real CLI and native
+processes, the production `FileAuthorityStore`, and process death at persistence
+boundaries. [Installed-package E2E](installed.py) repeats the public lifecycle
+outside the checkout for both wheel and sdist. [Negative controls](mutants.py)
+deliberately remove correctness checks in disposable source copies.
+
+These checks do not change the nine pending `s2c2.*` ladder declarations above.
+Sustained production parity, the migration/growth gates, and promotion remain
+separate obligations. A bounded qualification result reports
+`sustained_parity_verdict=not_evaluated`.
+
+Future ladder rows must use the actual product interfaces:
+
+- `authority-shadow drain` for receipt-verified replay and cursor recovery;
+- `coordination-shadow inspect / qualify / read-candidate` for comparison,
+  bounded historical qualification, and a qualified read from that same head;
+- `coordination-shadow rollback` with an exact revision or unfinished bootstrap
+  operation selector, followed by explicit rebootstrap;
+- the stable management lock outside the goal outbox for scheduling a management
+  boundary. Rollback moves the entire goal outbox, so a lock inside that directory
+  cannot coordinate it. The tests retain scheduling-only barriers around real
+  persistence calls rather than depending on the retired drain lock.
+
+There is no `verify` command or generic `reset`. The independent v0 observation
+store and its runtime-root override behavior retain their original scope; its
+historical evidence is not an outbox qualification receipt.