fix(model-governance): persist policy floor source of truth

config/model-governance/model-governance.staging.json

@@ -45,7 +45,7 @@
       "daily_token_budget": 250000,
       "denied_providers": [],
       "fallback_limit": 1,
-      "max_input_tokens_per_request": 16000,
+      "max_input_tokens_per_request": 16384,
       "max_output_tokens_per_request": 512,
       "project_id": null,
       "required_capabilities": [
@@ -63,7 +63,7 @@
       "daily_token_budget": 250000,
       "denied_providers": [],
       "fallback_limit": 1,
-      "max_input_tokens_per_request": 16000,
+      "max_input_tokens_per_request": 16384,
       "max_output_tokens_per_request": 512,
       "project_id": "staging",
       "required_capabilities": [

docs/superpowers/specs/2026-08-18-model-governance-policy-profile-floor-compatibility.md

@@ -0,0 +1,109 @@
+# Model Governance Policy / Profile Floor Compatibility
+
+## Problem
+
+The root-owned OpenClaw runtime profile reserves a minimum of 16,384 input tokens so Model
+Governance can cover OpenClaw-injected workspace, skills and system context before provider
+execution.
+
+The staging scope policies still limited input to 16,000 tokens per request. Route selection
+therefore failed before reservation with:
+
+```text
+MODEL_INPUT_TOKEN_LIMIT_DENY
+```
+
+This was a configuration invariant violation:
+
+```text
+scope max_input_tokens_per_request
+<
+root-owned profile input_reservation_floor_tokens
+```
+
+The registered OpenClaw model itself supports 32,768 input tokens, so the model capability was not
+the blocker.
+
+## Invariant
+
+For every Model Governance scope that allows a model/runtime whose root-owned profile declares an
+input reservation floor:
+
+```text
+scope.max_input_tokens_per_request
+>=
+profile.model_governance.input_reservation_floor_tokens
+```
+
+The profile floor remains root-owned. Task payloads cannot alter it.
+
+## Staging policy
+
+The two staging policies that allow `openai/deepseek-v4-flash` are updated from:
+
+```text
+16000
+```
+
+to:
+
+```text
+16384
+```
+
+This is the minimum compatible value. It intentionally remains below the model's 32,768-token
+technical input ceiling.
+
+The daily token budget is unchanged.
+
+## Governance path
+
+The live SQLite policy must be changed only through the repository bootstrap and
+`aios_model_governance_admin.py`.
+
+The admin reconciler:
+
+- reads the governed repository config;
+- uses current policy revision as CAS;
+- calls `ModelGovernanceStore.set_scope_policy`;
+- appends normal Model Governance audit evidence;
+- verifies the audit chain.
+
+Direct SQL updates remain forbidden.
+
+## Exhausted OpenClaw recovery
+
+The fixed OpenClaw live Task has exhausted its normal three attempts.
+
+After the governed policy candidate is officially verified and the live policy is reconciled:
+
+```text
+attempt 3 VERIFYING/FAILED result
+→ official NEEDS_REWORK
+→ Task becomes FAILED 3/3
+→ official bounded retry extension 3 → 4
+→ same Task / same Job returns QUEUED
+→ persistent OpenClaw worker auto-claims attempt 4
+```
+
+AR1 attempt-scoped external request identity ensures attempt 4 receives a new external request and a
+new Model Governance reservation.
+
+No historical reservation is deleted or rewritten.
+
+## Acceptance
+
+Attempt 4 must prove:
+
+```text
+scope max_input_tokens_per_request=16384
+profile input_reservation_floor_tokens=16384
+reserved_estimated_input_tokens>=16384
+actual input <= reserved input
+settlement=SETTLED
+usage_source=runtime
+audit_chain_valid=true
+```
+
+If attempt 4 exposes a new failure, stop. No additional retry extension is authorized by this
+repair.

tests/test_model_governance_runtime_config.py

@@ -65,3 +65,35 @@ def test_staging_bootstrap_has_explicit_tenant_live_policy():
     }
     assert ("legacy",None,None) in keys
     assert ("tenant","model-governance-live","staging") in keys
+
+
+def test_staging_policy_input_limit_covers_root_owned_openclaw_reservation_floor():
+    config=json.loads(
+        (ROOT/"config/model-governance/model-governance.staging.json").read_text()
+    )
+    profile=json.loads(
+        (ROOT/"config/external-agent-runtime.d/openclaw.profile.json").read_text()
+    )
+    floor=int(profile["model_governance"]["input_reservation_floor_tokens"])
+    assert floor==16384
+    openclaw_model="openai/deepseek-v4-flash"
+    relevant=[
+        policy for policy in config["policies"]
+        if openclaw_model in policy["allowed_models"]
+    ]
+    assert relevant
+    for policy in relevant:
+        assert int(policy["max_input_tokens_per_request"]) >= floor
+
+
+def test_staging_openclaw_scope_limit_is_least_privilege_floor_not_model_ceiling():
+    config=json.loads(
+        (ROOT/"config/model-governance/model-governance.staging.json").read_text()
+    )
+    models={item["model_id"]:item for item in config["models"]}
+    model=models["openai/deepseek-v4-flash"]
+    assert int(model["max_input_tokens"])==32768
+    for policy in config["policies"]:
+        if "openai/deepseek-v4-flash" in policy["allowed_models"]:
+            assert int(policy["max_input_tokens_per_request"])==16384
+            assert int(policy["max_input_tokens_per_request"]) < int(model["max_input_tokens"])