Evaluate nightly allocator API for event loop

.github/workflows/anyio.yml

@@ -69,9 +69,9 @@ jobs:
           cache-suffix: anyio-${{ matrix.os_name }}-py${{ matrix.python_version }}
 
       - name: Set up Rust
-        uses: dtolnay/rust-toolchain@stable
+        uses: dtolnay/rust-toolchain@master
         with:
-          toolchain: "1.98.1"
+          toolchain: "nightly-2026-09-25"
 
       - name: Cache Cargo registry
         uses: Swatinem/rust-cache@v2

.github/workflows/tests.yml

@@ -48,9 +48,9 @@ jobs:
           python-version: "3.13"
 
       - name: Set up Rust
-        uses: dtolnay/rust-toolchain@stable
+        uses: dtolnay/rust-toolchain@master
         with:
-          toolchain: "1.98.1"
+          toolchain: "nightly-2026-09-25"
 
       - name: Cache Cargo registry
         uses: Swatinem/rust-cache@v2
@@ -80,9 +80,9 @@ jobs:
           python-version: "3.13"
 
       - name: Set up Rust
-        uses: dtolnay/rust-toolchain@stable
+        uses: dtolnay/rust-toolchain@master
         with:
-          toolchain: "1.98.1"
+          toolchain: "nightly-2026-09-25"
           components: clippy
 
       - name: Cache Cargo registry
@@ -91,6 +91,30 @@ jobs:
       - name: Run Clippy
         run: cargo clippy --all-targets --all-features --locked -- -D warnings
 
+  allocator-miri:
+    name: allocator miri (strict provenance)
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - name: Check out repository
+        uses: actions/checkout@v6
+        with:
+          persist-credentials: false
+
+      - name: Set up Rust
+        uses: dtolnay/rust-toolchain@master
+        with:
+          toolchain: "nightly-2026-09-25"
+          components: miri,rust-src
+
+      - name: Cache Cargo registry
+        uses: Swatinem/rust-cache@v2
+
+      - name: Run allocator tests under Miri
+        env:
+          MIRIFLAGS: -Zmiri-strict-provenance
+        run: cargo miri test --manifest-path tools/vibeio-check/Cargo.toml --features scheduler-batch-cache batch_
+
   type-check:
     name: pyright
     runs-on: ubuntu-latest
@@ -165,9 +189,9 @@ jobs:
           cache-suffix: anyio
 
       - name: Set up Rust
-        uses: dtolnay/rust-toolchain@stable
+        uses: dtolnay/rust-toolchain@master
         with:
-          toolchain: "1.98.1"
+          toolchain: "nightly-2026-09-25"
 
       - name: Cache Cargo registry
         uses: Swatinem/rust-cache@v2
@@ -283,9 +307,9 @@ jobs:
           cache-suffix: ${{ matrix.os_name }}-py${{ matrix.python_version }}
 
       - name: Set up Rust
-        uses: dtolnay/rust-toolchain@stable
+        uses: dtolnay/rust-toolchain@master
         with:
-          toolchain: "1.98.1"
+          toolchain: "nightly-2026-09-25"
 
       - name: Cache Cargo registry
         uses: Swatinem/rust-cache@v2

.github/workflows/vibeio-native.yml

@@ -47,9 +47,9 @@ jobs:
           persist-credentials: false
 
       - name: Set up Rust
-        uses: dtolnay/rust-toolchain@stable
+        uses: dtolnay/rust-toolchain@master
         with:
-          toolchain: "1.98.1"
+          toolchain: "nightly-2026-09-25"
           components: clippy, rustfmt
 
       - name: Cache runtime harness build

.github/workflows/wheels.yml

@@ -64,9 +64,9 @@ jobs:
         uses: astral-sh/setup-uv@08807647e7069bb48b6ef5acd8ec9567f424441b
 
       - name: Set up Rust
-        uses: dtolnay/rust-toolchain@stable
+        uses: dtolnay/rust-toolchain@master
         with:
-          toolchain: "1.98.1"
+          toolchain: "nightly-2026-09-25"
 
       - name: Install LLVM tools for PGO
         if: github.event_name == 'workflow_dispatch' && inputs.pgo
@@ -114,9 +114,9 @@ jobs:
         uses: astral-sh/setup-uv@08807647e7069bb48b6ef5acd8ec9567f424441b
 
       - name: Set up Rust
-        uses: dtolnay/rust-toolchain@stable
+        uses: dtolnay/rust-toolchain@master
         with:
-          toolchain: "1.98.1"
+          toolchain: "nightly-2026-09-25"
 
       - name: Cache Cargo registry and target
         uses: Swatinem/rust-cache@v2

Cargo.toml

@@ -60,6 +60,8 @@ windows-sys = { version = "0.61", features = ["Wdk_Foundation", "Wdk_Storage_Fil
 
 [features]
 default = []
+# Opt-in: faster scheduler turns, but measured TCP workloads can regress.
+scheduler-batch-cache = []
 blocking-default = ["dep:rusty_pool"]
 # Embedded-runtime modules are opt-in; production wheels retain the default
 # networking/timer build unless a feature is explicitly requested.

benches/compare_event_loops.py

@@ -3,6 +3,7 @@
 
 import argparse
 import asyncio
+import contextvars
 import ctypes
 import gc
 import importlib
@@ -373,6 +374,37 @@ async def tiny_task() -> None:
     )
 
 
+async def bench_task_options(
+    loop_name: str, iterations: int, batch_size: int
+) -> ChildResult:
+    """Exercise named, explicit-context Task construction (Python 3.11+)."""
+
+    async def tiny_task() -> None:
+        await asyncio.sleep(0)
+
+    loop = asyncio.get_running_loop()
+    context = contextvars.copy_context()
+    start = time.perf_counter()
+    remaining = iterations
+    while remaining > 0:
+        current_batch = min(batch_size, remaining)
+        tasks = [
+            loop.create_task(tiny_task(), name="benchmark-task", context=context)
+            for _ in range(current_batch)
+        ]
+        await asyncio.gather(*tasks)
+        remaining -= current_batch
+    return ChildResult(
+        loop_name,
+        "task_options",
+        time.perf_counter() - start,
+        iterations,
+        baseline_rss_bytes=0,
+        peak_rss_bytes=0,
+        peak_rss_delta_bytes=0,
+    )
+
+
 async def maybe_wait_closed(writer: asyncio.StreamWriter) -> None:
     wait_closed = getattr(writer, "wait_closed", None)
     if wait_closed is None:

docs/allocator-batch-results.md

@@ -0,0 +1,146 @@
+# Scheduler batch allocation reuse
+
+Measured on 2026-09-25. This change uses the stabilized allocator API available
+in the pinned `nightly-2026-09-25` compiler, without an `allocator_ext` feature
+gate. It does **not** claim compatibility with an older stable-channel compiler.
+
+The cache is **opt-in** through `scheduler-batch-cache`, not enabled in default
+builds or wheels. Its scheduler-turn improvement comes with a confirmed TCP
+slowdown, so the general-purpose event loop retains its original allocation
+and drain path.
+
+## What changes
+
+`Runtime::block_on` previously allocated and freed its ready-task vector on
+every entry. With the feature enabled, the runtime owns a single-slot, exact-layout
+`System` allocation cache, used through `Allocator` and `Vec::with_capacity_in`.
+Successive calls reuse that storage; batches within a call keep the vector's
+capacity. Tests verify one system allocation across repeated runtime turns,
+including root-future unwinding and reuse afterward.
+
+The usual retained allocation is 2 KiB on 64-bit systems, with a hard 16 KiB
+ceiling per runtime. Runtime destruction frees it. This reduces allocator
+traffic, not the live size of tasks or a guaranteed amount of process RSS.
+No mutex, allocator `Arc` clone, size-class search, or public configuration is
+introduced. The cache sits outside `RuntimeInner` to preserve the hot shared
+scheduler state's layout.
+
+Allocator-aware `Vec::drain` is not in the stabilized subset. A private full
+drain uses stable vector primitives, retains the allocation, preserves FIFO
+order, and drops unconsumed tasks on unwind. Strict-provenance Miri tests cover
+partial consumption, destructor panics, zero-sized elements, forgotten drains,
+alignment, growth, zeroing, disjoint simultaneous allocations, and runtime
+panic cleanup. Kani's older compiler retains an ordinary-vector fallback;
+its proofs do not verify the custom allocator.
+
+The previous shared size-class allocator remains rejected. Early versions of
+this change also regressed busy turns when replacing the vector each batch or
+using optional task slots. Neither implementation is retained.
+
+## Native scheduler measurement
+
+Host: Intel Core i9-9900K, Linux x86-64; Rust 1.100.0-nightly
+(`f7575a9da`, LLVM 23.1.1). Identical `runtime_turns.rs` source and lockfile were
+built with the same compiler and release profile, against `d4f56f1` and the
+new implementation with `--features scheduler-batch-cache`. No other builds
+or tests ran during timing. CPU affinity
+was unrestricted; these are local results, not a cross-platform guarantee.
+
+Each comparison used 24 randomized, balanced baseline/candidate process pairs,
+10,000 warm-up turns per process, and paired bootstrap 95% confidence intervals.
+Negative percentages mean less elapsed time.
+
+| Workload | Turns per sample | Baseline median | Candidate median | Change (95% interval) |
+| --- | ---: | ---: | ---: | ---: |
+| Idle scheduler entry/exit | 7,000,000 | 0.76260 s | 0.64408 s | -15.41% [-15.87%, -14.88%] |
+| 64 continuously ready tasks | 500,000 | 1.12497 s | 1.12506 s | +0.30% [-0.48%, +1.14%] |
+
+The idle-turn gain passed the declared 1% timing threshold; the busy result is
+inconclusive and within the 3% regression budget at this confidence level.
+These measurements isolate scheduler turns, not Python or network throughput.
+
+Reproduce with `scripts/compare_runtime_turns.py` as described in
+[Development](development.md). Use `--iterations 7000000 --tasks 0 --seed 999`
+for idle turns and `--iterations 500000 --tasks 64 --seed 1000` for busy turns.
+Both use `--blocks 24` and separate output directories.
+
+Local raw plans, binary hashes, paired samples, and reports are in
+`target/allocator-batch-optin-idle` and `target/allocator-batch-optin-active`.
+These ignored directories are local experiment artifacts, not checked-in data.
+
+## Python application workloads
+
+The six-workload holdout used CPython 3.14.7, 24 paired process blocks,
+seed 997, and `tcp_streams` as the predeclared primary. Compiler versions,
+non-PGO flags, benchmark source, and artifact hashes were checked by the
+hot-path harness. Each interval below uses 99.583% confidence (Bonferroni
+adjustment across six timing and six RSS metrics).
+
+| Workload | Time change (interval) | Peak RSS change (interval) |
+| --- | ---: | ---: |
+| Callbacks | -0.14% [-0.70%, +0.36%] | +0.00% [-0.01%, +0.02%] |
+| Tasks | -0.64% [-1.50%, +0.14%] | -0.19% [-0.41%, +0.01%] |
+| TCP streams | +2.10% [+0.64%, +3.56%] | +0.00% [-0.23%, +0.23%] |
+| HTTP keep-alive | -0.30% [-6.67%, +3.78%] | +0.01% [-0.28%, +0.32%] |
+| Mixed streams | -0.71% [-2.98%, +1.33%] | +0.03% [-0.29%, +0.38%] |
+| Bulk transfer | -0.43% [-2.01%, +0.82%] | -0.00% [-0.02%, +0.01%] |
+
+The application-wide gate is **inconclusive**, not passed: the primary did
+not improve, TCP showed a small slowdown, and the TCP/HTTP upper timing bounds
+exceed the 3% budget. Peak RSS stayed well within budget, but there is no
+established application-wide memory reduction or throughput improvement.
+
+An independent 40-pair TCP confirmation (seed 998) found **+2.75%** elapsed time,
+with a 97.5% interval of **[+1.79%, +3.85%]**. RSS changed +0.04%
+[-0.14%, +0.24%]. This confirmed the tradeoff and is why the cache is opt-in.
+Raw confirmation data is in `target/allocator-batch-final-tcp-confirmation`.
+
+The exact baseline is `d4f56f1`; the local cache-enabled measurement snapshot is
+`a87092ec214b6daa9288818109f58cb6034846d3`. It predates the final feature guard:
+these application results evaluate the cache integration, not the final default
+build (which uses the original path). Artifact directories
+are `target/allocator-batch-baseline-final` and `target/allocator-batch-final`;
+the complete plan, samples, manifests and report are under
+`target/allocator-batch-final-holdout`.
+
+The following command records the cache-enabled snapshot comparison. To repeat
+the experiment on later revisions, use a cache-enabled candidate build; an
+ordinary default build no longer enables the cache. Use fresh output directories.
+
+```bash
+.venv/bin/python scripts/hotpath_lab.py compare \
+  --baseline target/allocator-batch-baseline-final \
+  --candidate target/allocator-batch-final \
+  --primary tcp_streams --suite holdout --blocks 24 --seed 997 \
+  --out target/allocator-batch-final-holdout
+```
+
+## Correctness checks
+
+- Root Rust suites: 307 passed with default features; 407 with all features.
+- Embedded runtime all-feature suite: 283 passed; 16 doctests passed.
+- Strict-provenance Miri: 10 allocator/drain/runtime regression tests passed.
+- Kani: all 31 `merge_` harnesses passed, using the ordinary-vector fallback.
+- Clippy: both crates, all targets/features, warnings denied.
+- Python: 192 passed and 2 skipped on **each** isolated extension; tooling,
+  stress, and slow-network markers were excluded (73 deselected).
+- Final default development install: the same 192 Python tests passed, with
+  2 skipped and 73 deselected.
+- Repository tooling: 67 passed. The new comparison runner also passes Ruff.
+
+Relevant commands:
+
+```bash
+python3 scripts/run_rust_tests.py --all-features
+cargo test --manifest-path tools/vibeio-check/Cargo.toml --all-features --locked
+MIRIFLAGS=-Zmiri-strict-provenance cargo miri test \
+  --manifest-path tools/vibeio-check/Cargo.toml --features scheduler-batch-cache batch_
+cargo kani --harness merge_ -j 2 --output-format terse
+PYTEST_ADDOPTS="-m 'not tooling and not stress and not slow_network'" \
+  .venv/bin/python scripts/hotpath_lab.py test \
+  --artifact target/allocator-batch-final --python-child
+```
+
+The Python command was also run against `target/allocator-batch-baseline-final`.
+These checks do not replace cross-platform CI or the excluded stress/network
+suites.