feat(storage): reclaim disk safely during live queries

.github/workflows/tests.yaml

@@ -23,7 +23,7 @@ jobs:
     # self-hosted host. Everything else (in-repo branches, push, dispatch) uses the faster
     # self-hosted runner with a warm cargo cache. Runner groups can't gate by trigger, so the
     # split has to live here in the workflow — not in the approval setting.
-    runs-on: ${{ (github.event_name == 'pull_request' && github.event.pull_request.head.repo.fork) && fromJSON('["ubuntu-latest"]') || fromJSON('["self-hosted", "dev-server"]') }}
+    runs-on: ${{ (github.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name != github.repository) && fromJSON('["ubuntu-latest"]') || fromJSON('["self-hosted", "dev-server"]') }}
     if: github.event_name == 'push' || !github.event.pull_request.draft
     steps:
       - uses: actions/checkout@v4
@@ -50,7 +50,7 @@ jobs:
     name: Test
     # Fork PRs stay on GitHub-hosted (untrusted code); trusted events use self-hosted. See the
     # clippy job above for the rationale.
-    runs-on: ${{ (github.event_name == 'pull_request' && github.event.pull_request.head.repo.fork) && fromJSON('["ubuntu-latest"]') || fromJSON('["self-hosted", "dev-server"]') }}
+    runs-on: ${{ (github.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name != github.repository) && fromJSON('["ubuntu-latest"]') || fromJSON('["self-hosted", "dev-server"]') }}
     if: github.event_name == 'push' || !github.event.pull_request.draft
     steps:
       - uses: actions/checkout@v4

crates/hotblocks-harness/README.md

@@ -23,6 +23,8 @@ is what makes the crash/restart and shutdown classes expressible at all.
 cargo test -p sqd-hotblocks-harness          # the harness's own unit tests (model, chain, simulator)
 cargo test -p sqd-hotblocks --test ct1_happy_path   # CT-1 — the Phase 0 exit criterion
 cargo test -p sqd-hotblocks --test ct9_source_faults
+# Explicit endurance lane (ignored by default; includes the reclaim convergence wait):
+cargo test -p sqd-hotblocks --test ct7_stall_and_churn ct7_churn_soak -- --ignored
 ```
 
 The CT tests live in `crates/hotblocks/tests/` because only a test inside that package gets
@@ -127,11 +129,10 @@ Fixed in `crates/data-client/src/reqwest/lines.rs`; pinned by a unit test there
   normative CONFLICT recovery of 04 §7. What is missing is the scripts.
 - **CT-5 (error taxonomy)** — `Model::predict_query` returns the outcome class for any query;
   `Client::query` already classifies every response. What is missing is the request matrix.
-- **CT-7 (soak/space)** — needs `OB-6`-style space accounting; the retention model transition
-  (`Model::retain`) is implemented but the comparator compares the first block *exactly*, which
-  is wrong once retention starts trimming: the service trims whole chunks, so its window may be
-  larger than the model's (legal under RS-3/RS-4, `P-RETENTION-SLACK`). Give the comparator that
-  tolerance before writing retention tests.
+- **CT-7 (soak)** — the ignored S4 runner in `ct7_stall_and_churn` drives API-controlled
+  moving-window retention and samples `total-sst-files-size`, `estimate-live-data-size`, and
+  memtable bytes through the existing RocksDB property endpoint. Prometheus OB-2/3/6 series are
+  intentionally deferred to a separate change.
 - **CT-9 (fuzz)** — `SimFaults` is the injection point.
 
 ## Known open questions

crates/hotblocks-harness/src/driver.rs

@@ -79,6 +79,7 @@ pub struct StatusData {
     pub finalized_head: Option<BlockRef>
 }
 
+#[derive(Clone)]
 pub struct Client {
     http: reqwest::Client,
     base: String,
@@ -178,6 +179,19 @@ impl Client {
         Ok(Metrics(out))
     }
 
+    /// Reads the service's existing intrinsic RocksDB diagnostic property surface.
+    pub async fn rocksdb_property(&self, column_family: &str, name: &str) -> Result<Option<u64>> {
+        let res = self
+            .http
+            .get(format!("{}/rocksdb/prop/{column_family}/{name}", self.base))
+            .send()
+            .await?;
+        if res.status() == reqwest::StatusCode::NOT_FOUND {
+            return Ok(None);
+        }
+        Ok(Some(res.error_for_status()?.text().await?.trim().parse()?))
+    }
+
     pub async fn query(&self, body: &Value) -> Result<Outcome> {
         self.query_at("stream", body).await
     }

crates/hotblocks-harness/src/lib.rs

@@ -44,6 +44,7 @@ pub mod driver;
 pub mod harness;
 pub mod model;
 pub mod sim;
+pub mod soak;
 pub mod sut;
 pub mod types;
 
@@ -56,5 +57,6 @@ pub use driver::{Client, Emitted, FollowStep, Follower, Outcome};
 pub use harness::{Harness, HarnessConfig};
 pub use model::{Finalize, ForkResolution, Model, Predicted};
 pub use sim::{Numbering, SourceSim};
+pub use soak::{ChurnSoakConfig, ChurnSoakReport, SpaceSample, StallProbe, StallReport};
 pub use sut::{DatasetSpec, Retention, Sut, SutConfig};
 pub use types::{Anchor, Block, BlockNumber, BlockRef};