diff --git a/planning/changes/2026-07-13.02-grow-integration-scenarios.md b/planning/changes/2026-07-13.02-grow-integration-scenarios.md
new file mode 100644
index 0000000..f2fbeca
--- /dev/null
+++ b/planning/changes/2026-07-13.02-grow-integration-scenarios.md
@@ -0,0 +1,103 @@
+---
+summary: Add three integration scenarios to the harness from PR #41 -- pod-level `dns`/`sysctls`/`extra_hosts` (the exact area that just broke), `secrets`/`configs`, and resource limits (`deploy.resources.limits` acceptance plus a value-checked `ulimits`) -- reusing the existing `run_pod` fixture unchanged.
+---
+
+# Design: grow the integration scenario set
+
+## Summary
+
+Three new scenarios in `tests/integration/`, each a single new test file using
+the `run_pod` fixture shipped in PR #41 (`planning/changes/2026-07-12.04-integration-harness.md`)
+unchanged. No harness, `conftest.py`, `pyproject.toml`, or CI changes. Every
+scenario is a behavioral probe (script exit code + stdout), matching the
+existing seed scenarios, for the same reason: the generated script
+self-destructs its pod on exit, so there is nothing to `podman exec`/`inspect`
+afterward.
+
+## Motivation
+
+The harness's first real CI run caught a genuine podman-compat bug in the
+pod-level `--add-host` emission (`planning/changes/2026-07-13.01-pod-level-add-host.md`)
+that every unit test missed. That is exactly the class of bug the harness
+exists to catch, and the areas most likely to hide another one are the ones
+that talk to podman in ways a string-comparison unit test cannot verify:
+pod-wide `dns`/`sysctls`/`extra_hosts` (only unit-tested even after the fix),
+`secrets`/`configs` (podman secret lifecycle), and resource limits (podman
+flag acceptance). This round adds one scenario per area.
+
+## Design
+
+**Scenario 1 -- pod-level options** (`test_pod_level_options.py`): one
+service declares `dns: ["9.9.9.9"]`, `sysctls: {"net.core.somaxconn":
+"1024"}`, and `extra_hosts: {"external-svc": "10.0.0.9"}`; its command chains
+three checks with `&&`:
+```
+grep 9.9.9.9 /etc/resolv.conf && grep 1024 /proc/sys/net/core/somaxconn && grep external-svc:10.0.0.9 /etc/hosts
+```
+No dependency ordering needed -- one container reading its own environment.
+`net.core.somaxconn` is network-namespaced and known rootless-safe (podman's
+own docs use it as the canonical rootless sysctl example); `dns`/`extra_hosts`
+have no such restriction.
+
+**Scenario 2 -- secrets/configs** (`test_secrets_and_configs.py`): one
+service references one `secrets:` entry (`file:` source) and one `configs:`
+entry (`content:` source, the one source `configs` supports that `secrets`
+doesn't -- exercising both source-kind branches in `compose2pod/stores.py` in
+one pass). Command: `cat /run/secrets/mysecret && cat /myconfig`, asserting
+both contents appear in stdout. Proves the full lifecycle: `podman secret
+create`, the `--secret source=...,target=...` mount at both default target
+shapes, and that `podman secret rm` teardown doesn't interfere with the run.
+
+**Scenario 3 -- resource limits** (`test_resource_limits.py`): one service
+sets `deploy: {resources: {limits: {memory: "128m", cpus: "0.5", pids:
+100}}}` and `ulimits: {nofile: 1024}`. Command:
+`[ "$(ulimit -n)" = 1024 ] && echo ok`. Two distinct verification strategies,
+deliberately different in strength:
+- `memory`/`cpus`/`pids` (-> `--memory`/`--cpus`/`--pids-limit`): **acceptance-only**
+  -- if podman rejected any of these flags (the same failure class as the
+  add-host bug), the script would fail to start and `run.returncode` would be
+  nonzero. No value is checked from inside the container.
+- `ulimits` (-> `--ulimit`): **value-checked** via the `ulimit -n` shell
+  builtin, cgroup-independent and reliable rootless or rootful on any runner.
+
+The split is deliberate, not an oversight: verifying the exact `--memory`/
+`--pids-limit` value would require reading `/sys/fs/cgroup/memory.max` /
+`pids.max` inside the container, which assumes a cgroup v2 unified hierarchy
+on the CI runner. That assumption could make the scenario flake for reasons
+that have nothing to do with compose2pod's correctness. Acceptance-only
+targets the actual risk class this harness exists to catch (podman refusing
+a flag combination outright) at zero environment-fragility cost.
+
+## Non-goals
+
+- `dns_search`/`dns_opt` -- same code family as `dns` (`_DNS_KEYS` in
+  `compose2pod/pod.py`), lower marginal value; not a new code path.
+- Cgroup-based value verification for `memory`/`pids` -- environment-fragile
+  (cgroup v1 vs v2), deferred; see Design above.
+- `deploy.resources.reservations` -- a scheduling hint (`--memory-reservation`),
+  not a hard limit; podman does not expose a way to observe it from inside a
+  container, so there is nothing behavioral to assert.
+- Multi-service secret/config sharing -- a single service is enough to prove
+  the `podman secret create` + mount mechanism; sharing one secret across
+  services is unit-tested already (`tests/test_stores.py`).
+
+## Testing
+
+`uv run --no-sync pytest -m integration --collect-only -q` collects 6 tests
+(the 3 existing seed scenarios plus these 3). `just test-ci` stays at 100%
+with all 6 deselected. `just lint-ci` and `just check-planning` clean. The
+real signal is the `integration` CI job going green against real podman on
+`ubuntu-latest`, same as the seed scenarios.
+
+## Risk
+
+- **Rootless sysctl restriction** (low x med): if `net.core.somaxconn` turns
+  out not to be settable on the GitHub Actions rootless podman build despite
+  being the documented canonical example, the scenario fails for an
+  environment reason unrelated to compose2pod. Mitigated by picking the
+  specific sysctl podman's own docs use for this exact purpose.
+- **Acceptance-only under-verifies `deploy.resources.limits`** (low x low): a
+  regression that silently drops the `--memory` value (but keeps the flag
+  present with a garbage value) would not be caught. Accepted per the Design
+  section's fragility trade-off; a future round can add cgroup-file
+  verification if this trade-off turns out wrong in practice.
diff --git a/tests/integration/test_pod_level_options.py b/tests/integration/test_pod_level_options.py
new file mode 100644
index 0000000..4587866
--- /dev/null
+++ b/tests/integration/test_pod_level_options.py
@@ -0,0 +1,34 @@
+"""Pod-level dns/sysctls/extra_hosts merge onto podman pod create -- the area that broke first."""
+
+from collections.abc import Callable
+
+from tests.integration.conftest import PodRun
+
+
+def test_pod_level_options_land_on_the_pod(run_pod: Callable[..., PodRun]) -> None:
+    compose = {
+        "services": {
+            "app": {
+                "image": "busybox:1.36",
+                "dns": ["9.9.9.9"],
+                "sysctls": {"net.core.somaxconn": "1024"},
+                "extra_hosts": {"external-svc": "10.0.0.9"},
+                "command": [
+                    "sh",
+                    "-c",
+                    "grep 9.9.9.9 /etc/resolv.conf "
+                    "&& grep 1024 /proc/sys/net/core/somaxconn "
+                    "&& grep 10.0.0.9 /etc/hosts "
+                    "&& grep external-svc /etc/hosts",
+                ],
+            },
+        },
+    }
+    run = run_pod(compose, target="app")
+    # Exit 0 proves all three landed on the pod (podman pod create), not the
+    # container -- exactly the merge that regressed in the add-host bug.
+    # /etc/hosts is "ip<whitespace>hostname" (podman's --add-host FLAG syntax
+    # is "host:ip", but the file it writes is the other order, unquoted) --
+    # checked as two independent substrings rather than one combined pattern
+    # to stay robust to the exact separator.
+    assert run.returncode == 0, run.stderr
diff --git a/tests/integration/test_resource_limits.py b/tests/integration/test_resource_limits.py
new file mode 100644
index 0000000..9f9ed1c
--- /dev/null
+++ b/tests/integration/test_resource_limits.py
@@ -0,0 +1,32 @@
+"""Resource limits: deploy.resources.limits is acceptance-only, ulimits is value-checked.
+
+Verifying the exact --memory/--pids-limit value would require reading cgroup files
+inside the container, which assumes a cgroup v2 unified hierarchy on the CI runner --
+an assumption that could make this scenario flake for reasons unrelated to compose2pod's
+correctness. Acceptance-only (the script simply runs to completion) targets the actual
+risk class this harness exists to catch: podman refusing a flag combination outright,
+the same failure mode as the add-host bug. ulimits has no such fragility -- `ulimit -n`
+is a shell builtin, independent of cgroups -- so its value is checked directly.
+"""
+
+from collections.abc import Callable
+
+from tests.integration.conftest import PodRun
+
+
+def test_resource_limits_are_accepted_and_ulimit_is_applied(run_pod: Callable[..., PodRun]) -> None:
+    compose = {
+        "services": {
+            "app": {
+                "image": "busybox:1.36",
+                "deploy": {"resources": {"limits": {"memory": "128m", "cpus": "0.5", "pids": 100}}},
+                "ulimits": {"nofile": 1024},
+                "command": ["sh", "-c", '[ "$(ulimit -n)" = 1024 ] && echo ulimit-ok'],
+            },
+        },
+    }
+    run = run_pod(compose, target="app")
+    # Exit 0 proves podman accepted --memory/--cpus/--pids-limit (or the run would
+    # fail to start) AND that --ulimit nofile=1024 took effect inside the container.
+    assert run.returncode == 0, run.stderr
+    assert "ulimit-ok" in run.stdout
diff --git a/tests/integration/test_secrets_and_configs.py b/tests/integration/test_secrets_and_configs.py
new file mode 100644
index 0000000..92fc9af
--- /dev/null
+++ b/tests/integration/test_secrets_and_configs.py
@@ -0,0 +1,30 @@
+"""Secrets and configs: both compile to podman secret create, mounted at different paths."""
+
+from collections.abc import Callable
+from pathlib import Path
+
+from tests.integration.conftest import PodRun
+
+
+def test_secret_and_config_are_mounted(run_pod: Callable[..., PodRun], tmp_path: Path) -> None:
+    (tmp_path / "secret.txt").write_text("shh-secret-99\n")
+    compose = {
+        "services": {
+            "app": {
+                "image": "busybox:1.36",
+                "secrets": ["mysecret"],
+                "configs": ["myconfig"],
+                "command": ["sh", "-c", "cat /run/secrets/mysecret && cat /myconfig"],
+            },
+        },
+        "secrets": {"mysecret": {"file": "secret.txt"}},
+        "configs": {"myconfig": {"content": "config-value-7\n"}},
+    }
+    run = run_pod(compose, target="app", project_dir=tmp_path)
+    # Exit 0 + both contents in stdout proves: podman secret create (file source
+    # for the secret, content source for the config), the --secret mount at both
+    # default target shapes (/run/secrets/<name> and /<name>), and that secret rm
+    # teardown doesn't interfere with the run.
+    assert run.returncode == 0, run.stderr
+    assert "shh-secret-99" in run.stdout
+    assert "config-value-7" in run.stdout