Make the cost-based chaos assertions reliable under parallel CI load

cost-avoidance and mixed-technology decide a parent by measured link cost
(etx * (1 + srtt_ms/100)). Under the local CI's 4-way-parallel chaos the fast
fiber link's measured srtt spikes from host scheduling and can exceed the
Bluetooth link's, flipping the choice and reddening a run that proves nothing
about the daemon. The old margin was ~0.4 cost units, which a ~24ms one-sided
scheduling blip closes.

Widen the Bluetooth delay to 150-250ms so the margin dwarfs any plausible
one-sided spike. The link still establishes well within the handshake budget
(a ~500ms RTT against a 30s stale-handshake window), so the losing candidate
is still a real, established peer rather than an absent one.

Add a netem mutation exclude_edges option and use it in mixed-technology to
pin the two links n08 chooses between, so a random degradation can't flip the
asserted comparison either. An unknown excluded edge is a hard error, since a
silent no-op would reintroduce the flakiness it exists to remove.

Validated by running both scenarios repeatedly under heavier-than-CI
concurrent load: the parent choice holds every time.

testing/chaos/scenarios/cost-avoidance.yaml

@@ -13,9 +13,21 @@
 #      n04  (test subject)
 #
 # n04 has two candidate parents at depth 1: n02 (via Bluetooth L2CAP)
-# and n03 (via fiber). Cost-based selection should pick n03 because:
+# and n03 (via fiber). Cost-based selection should pick n03 because
-#   effective_depth(n02) = 1 + ~1.4 (Bluetooth) = ~2.4
+# link_cost = etx * (1 + srtt_ms/100), so the far slower Bluetooth link
-#   effective_depth(n03) = 1 + ~1.01 (fiber) = ~2.01
+# costs far more:
+#   effective_depth(n02) = 1 + ~4.5 (Bluetooth, ~150-250ms) = ~5.5
+#   effective_depth(n03) = 1 + ~1.05 (fiber, ~1-5ms)        = ~2.05
+#
+# The Bluetooth delay was widened deliberately, and the size is set by load,
+# not by realism. At the original 15-40ms the cost margin was ~0.4, and even
+# 80-120ms was not enough: under the CI's parallel-chaos contention the
+# *measured* srtt of the fiber link spikes (host scheduling, not netem) and
+# was observed to exceed a 120ms Bluetooth base, flipping n04 to n02. At
+# 150-250ms the gap dwarfs any plausible one-sided fiber spike, so the
+# assertion below is a reliable verdict rather than a coin toss under load.
+# n04 still establishes the n02 link (it was parented to it when this flaked),
+# so the choice remains real rather than vacuous.
 #
 # Validation: tree snapshot shows n04's parent is n03.
 
@@ -44,11 +56,14 @@ netem:
     jitter_ms: [0, 1]
     loss_pct: [0, 0.5]
   link_policies:
-    # Bluetooth (L2CAP) link from n02 to n04
+    # Bluetooth (L2CAP) link from n02 to n04. Delay widened to 150-250ms so the
+    # fiber-vs-Bluetooth cost margin survives CI load spikes on the fiber
+    # probe's measured latency (see header). Well within handshake timeouts, so
+    # the n02 link still establishes and the choice stays real.
     - edges: ["n02-n04"]
       policy:
-        delay_ms: [15, 40]
+        delay_ms: [150, 250]
-        jitter_ms: [5, 15]
+        jitter_ms: [15, 30]
         loss_pct: [2, 8]
 
 link_flaps:

testing/chaos/scenarios/mixed-technology.yaml

@@ -44,8 +44,8 @@
 #
 # Netem mutation shifts fiber-only links between normal and degraded.
 #
-# NEITHER TEST SUBJECT IS ASSERTED YET, and the reason has been found and
+# n08 IS now asserted (below); n06 is not, and both reasons trace to the root,
-# fixed at the root rather than papered over here.
+# which was found and fixed rather than papered over here.
 #
 # The diagram above puts n01 at the top, and both test subjects only make
 # sense in a tree rooted there. The mesh does not root by topology: it roots
@@ -64,13 +64,15 @@
 #
 # `topology.pin_root` (default true) now assigns identities in NodeAddr order,
 # so n01 holds the smallest and the diagram describes the tree that forms.
-# Both criteria should therefore become reachable and assertable.
+# Under that root n08 takes fiber n03 as its subject assertion (below) now
+# checks. n06 stays unasserted: its "fiber n03" was a WiFi link (see the
+# corrected note above), so it was never a founded criterion. Tracked in
+# ISSUE-2026-0085.
 #
-# STILL TO DO: run this under the pinned root, confirm n06 and n08 both take
+# The n08 assertion is made reliable, not merely reachable: the Bluetooth
-# n03, and encode `tree_parents` for them. Until then the baseline block below
+# contrast is widened (netem below) and the two links n08 chooses between are
-# is a floor and not a substitute — it holds the mesh to forming a tree and
+# excluded from the netem mutation, so neither host load nor a random
-# carrying traffic, and nothing here yet checks that the tree is the one these
+# degradation can flip the asserted comparison.
-# comments describe. Tracked in ISSUE-2026-0085.
 
 scenario:
   name: "mixed-technology"
@@ -105,11 +107,15 @@ netem:
     jitter_ms: [0, 1]
     loss_pct: [0, 0.5]
   link_policies:
-    # Bluetooth (L2CAP) links
+    # Bluetooth (L2CAP) links. Delay widened to 150-250ms so the
+    # fiber-vs-Bluetooth cost margin on n08's choice (fiber n03 vs Bluetooth
+    # n04) dwarfs any load-induced srtt spike on the fiber probe, matching
+    # cost-avoidance. Combined with excluding n03-n08/n04-n08 from the mutation
+    # (below), n08's asserted comparison is robust to both load and churn.
     - edges: ["n02-n06", "n04-n08"]
       policy:
-        delay_ms: [15, 40]
+        delay_ms: [150, 250]
-        jitter_ms: [5, 15]
+        jitter_ms: [15, 30]
         loss_pct: [2, 8]
     # WiFi links (moderate latency, low loss)
     - edges: ["n03-n06", "n08-n10"]
@@ -120,6 +126,12 @@ netem:
   mutation:
     interval_secs: {min: 30, max: 60}
     fraction: 0.2
+    # Never mutate the two links n08's tree_parents assertion depends on: the
+    # degraded policy (50-100ms) could otherwise slow the fiber n03-n08 link
+    # enough to flip n08 to n04 independently of the widened Bluetooth contrast,
+    # making the assertion a coin toss. The mutation still churns the other ten
+    # links, so the dynamic behaviour it exists to exercise is unchanged.
+    exclude_edges: ["n03-n08", "n04-n08"]
     policies:
       normal:
         delay_ms: [1, 10]

testing/chaos/sim/netem.py

@@ -144,6 +144,27 @@ class NetemManager:
                 len(self._edge_overrides),
             )
 
+        # Edges the periodic mutation must never touch (canonical "nXX-nYY").
+        # Unlike a link_policy typo, which merely fails to shape a link, a typo
+        # here would silently leave an asserted link unprotected and reintroduce
+        # the exact flakiness the exclusion exists to remove — so an unknown
+        # edge is a hard error, not a warning.
+        self._mutation_exclude: set[str] = set()
+        for edge_str in config.mutation.exclude_edges:
+            canonical = "-".join(sorted(edge_str.split("-")))
+            if canonical not in topo_edge_strs:
+                raise ValueError(
+                    f"netem.mutation.exclude_edges references {edge_str!r}, "
+                    "which is not an edge in the topology"
+                )
+            self._mutation_exclude.add(canonical)
+        if self._mutation_exclude:
+            log.info(
+                "Mutation excludes %d edge(s): %s",
+                len(self._mutation_exclude),
+                ", ".join(sorted(self._mutation_exclude)),
+            )
+
     def _htb_rate(self, node_id: str, peer_id: str) -> str:
         """Return the HTB rate string for a link direction."""
         rate = self._edge_rates.get((node_id, peer_id), 0)
@@ -370,10 +391,12 @@ class NetemManager:
         if not self.config.mutation.policies:
             return
 
-        # Only consider edges where both endpoints are up
+        # Only consider edges where both endpoints are up, and never the
+        # explicitly excluded ones (links an assertion depends on).
         live_edges = [
             (a, b) for a, b in self.topology.edges
             if a not in self.down_nodes and b not in self.down_nodes
+            and "-".join(sorted([a, b])) not in self._mutation_exclude
         ]
         if not live_edges:
             return

testing/chaos/sim/scenario.py

@@ -73,6 +73,11 @@ class NetemMutationConfig:
     interval_secs: Range = field(default_factory=lambda: Range(15, 30))
     fraction: float = 0.3
     policies: dict[str, NetemPolicy] = field(default_factory=dict)
+    # Edges the periodic mutation must never touch, "nXX-nYY" strings. Use it
+    # to pin the links a tree_parents assertion depends on so a random
+    # degradation cannot flip the very comparison being asserted. Validated
+    # against the topology in NetemManager — an unknown edge is a hard error.
+    exclude_edges: list[str] = field(default_factory=list)
 
 
 @dataclass
@@ -388,7 +393,7 @@ _SECTION_KEYS = {
     },
     "netem": {"enabled", "default_policy", "link_policies", "mutation"},
     "netem.link_policies[]": {"edges", "policy", "policy_name"},
-    "netem.mutation": {"interval_secs", "fraction", "policies"},
+    "netem.mutation": {"interval_secs", "fraction", "policies", "exclude_edges"},
     "link_flaps": {
         "enabled", "interval_secs", "max_down_links", "down_duration_secs",
         "protect_connectivity",
@@ -549,6 +554,7 @@ def load_scenario(path: str) -> Scenario:
             mc.get("interval_secs", {"min": 15, "max": 30}), "netem.mutation.interval_secs"
         )
         s.netem.mutation.fraction = float(mc.get("fraction", 0.3))
+        s.netem.mutation.exclude_edges = list(mc.get("exclude_edges", []))
         if "policies" in mc:
             s.netem.mutation.policies = {
                 name: _parse_netem_policy(pdata, f"netem.mutation.policies.{name}")