batch(0.69.5): EPIC #4249 patch train + #4313 f16 prefill — 0.69.5 RC

.github/actionlint.yaml

@@ -24,3 +24,5 @@ self-hosted-runner:
     - apple-silicon
     - m4
     - mini
+    # #4292 / infra#1042: the mini's dedicated repo-level release runner
+    - mini-builder

.github/workflows/cuda-nightly.yml

@@ -327,6 +327,42 @@ jobs:
           nice -n 19 cargo test -p aprender-serve --lib --release cb008_checker_controls
           APR_FALSIFY_MODEL="$model" nice -n 19 cargo test -p aprender-serve --features cuda \
             --release --test falsify_cb008_no_frozen_slots_2753 -- --nocapture
+
+      # aprender#4251: #4250's serve batched-prefill test. ci.yml's cuda-only step selects it, but
+      # there a missing Qwen3.5 file is a legitimate skip, and the runners do not carry one. So it
+      # passed without running. Here the model is resolved first, and a `SKIP:` line or a
+      # run that is not exactly `1 passed` is RED. `!cancelled()` so an earlier red falsifier
+      # cannot hide this one.
+      - name: aprender#4251 - every serve route prefills through the batched prefill (Qwen3.5)
+        if: ${{ !cancelled() && steps.decide.outputs.proceed == 'true' }}
+        env:
+          APR_QWEN35_MODEL: ${{ vars.APR_QWEN35_MODEL || '/home/noah/models/Qwen3.5-0.8B-Q4_K_M.gguf' }}
+        run: |
+          set -euo pipefail
+          if [ ! -f "$APR_QWEN35_MODEL" ]; then
+            echo "::error::#4251 UNMEASURABLE on ${{ matrix.name }} (provisioning, not a code defect): $APR_QWEN35_MODEL is absent"
+            exit 1
+          fi
+          echo "#4251 model: $APR_QWEN35_MODEL"
+          log="$RUNNER_TEMP/serve-batched-prefill.log"
+          rc=0
+          nice -n 19 cargo test -p aprender-serve --features cuda --lib --release \
+            api::openai_handlers::qwen35_serve_tests::gpu_every_serve_route_prefills_through_the_batched_prefill \
+            -- --exact --nocapture \
+            >"$log" 2>&1 || rc=$?
+          cat "$log"
+          if [ "$rc" -ne 0 ]; then
+            echo "::error::#4251 serve batched-prefill test FAILED (exit $rc)"
+            exit "$rc"
+          fi
+          if grep -q '^SKIP:' "$log"; then
+            echo "::error::#4251 the test SKIPPED on a CUDA runner with the model named: $(grep -m1 '^SKIP:' "$log")"
+            exit 1
+          fi
+          if ! grep -q 'test result: ok\. 1 passed' "$log"; then
+            echo "::error::#4251 the filter did not run exactly one test"
+            exit 1
+          fi
       # PP-26 / L0 CORRECTNESS: the batch-invariance witness (FALSIFY-CB-006,
       # #2753, §9 #1a on Blackwell). PP-LLAMA-001 §7.0 puts correctness at layer
       # L0 and says perf041 runs "nightly + release, missing marker = RED"; the

.github/workflows/install-script.yml

@@ -21,13 +21,15 @@ on:
     paths:
       - "scripts/install.sh"
       - "scripts/tests/install_test.sh"
+      - "scripts/tests/install_e2e_container.sh"
       - ".github/workflows/install-script.yml"
       - ".github/workflows/binary-release.yml"
       - ".github/workflows/nightly.yml"
   pull_request:
     paths:
       - "scripts/install.sh"
       - "scripts/tests/install_test.sh"
+      - "scripts/tests/install_e2e_container.sh"
       - ".github/workflows/install-script.yml"
       - ".github/workflows/binary-release.yml"
       - ".github/workflows/nightly.yml"
@@ -54,8 +56,13 @@ jobs:
 
       - name: "bashrs SEC/DET/IDEM (scripts/install.sh, scripts/tests/install_test.sh)"
         run: |
-          bashrs lint --level error scripts/install.sh
           bashrs lint --level error scripts/tests/install_test.sh
+          # #3546: install.sh and the container e2e are held to 0 errors AND 0 warnings, not just 0 errors.
+          for f in scripts/install.sh scripts/tests/install_e2e_container.sh; do
+            out="$(bashrs lint "$f" 2>&1 | sed 's/\x1b\[[0-9;]*m//g')" || true  # bashrs colours even a pipe
+            printf '%s\n' "$out" | grep -E 'Summary:'
+            printf '%s\n' "$out" | grep -qE 'Summary: 0 error\(s\), 0 warning\(s\)' || { printf '%s\n' "$out"; exit 1; }
+          done
 
       # Proves install_test.sh's own mutual-exclusivity row discriminates
       # before trusting the full run below — same discipline as
@@ -68,3 +75,24 @@ jobs:
       # it is the same contract a user's `curl | sh` exercises.
       - name: install_test.sh
         run: bash scripts/tests/install_test.sh
+
+  # #3546: the README one-liner end to end, per channel, nightly and on every install.sh change. See the script
+  # header for the five checks. RED today on `apr --version` for any release built without APR_GIT_SHA_OVERRIDE
+  # (it prints `+no-git`), which is the point: two rcs of one version must be told apart.
+  e2e-container:
+    name: curl | bash in a clean container (${{ matrix.channel }})
+    runs-on: [self-hosted, Linux, X64, clean-room, docker]
+    strategy:
+      fail-fast: false
+      matrix:
+        channel: [stable, rc]
+    steps:
+      - uses: actions/checkout@v7
+      - name: Preflight - docker
+        run: bash scripts/ci_self_hosted_preflight.sh --need docker
+      - name: version comparator case table
+        run: bash scripts/tests/install_e2e_container.sh --self-test
+      - name: curl | bash --channel ${{ matrix.channel }}
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: bash scripts/tests/install_e2e_container.sh ${{ matrix.channel }}

.github/workflows/qwen-story-daily.yml

@@ -84,9 +84,19 @@ jobs:
         shell: bash
         run: |
           set -euo pipefail
-          cargo install --path crates/apr-cli --force --features cuda 2>&1 | tail -5
+          # BUILD, never `cargo install`. This runner runs as the host user, so
+          # `cargo install` wrote the HOST's ~/.cargo/bin/apr - on gx10 that is
+          # the fleet-bins install dir, and every beat replaced the fleet's
+          # release binary with a CI dev build (measured 2026-09-25 01:30Z:
+          # v0.69.3 overwritten by `apr 0.69.0 (aa7c6ef0)`, .crates.toml naming
+          # this runner's _work checkout; fleet drift RED bytes-not-recorded).
+          # The story needs the binary on PATH, not installed; scripts/apr_bin.sh
+          # finds <target>/release/apr first and attributes it by dep-info.
+          cargo build --release -p apr-cli --bin apr --features cuda 2>&1 | tail -5
+          APR_BUILT_DIR="$(cargo metadata --format-version 1 --no-deps | jq -r .target_directory)/release"
+          [ -x "$APR_BUILT_DIR/apr" ] || { echo "::error::no apr at $APR_BUILT_DIR"; exit 1; }
 
-          # `cargo install` writes to ~/.cargo/bin. On this runner ~/.local/bin
+          # `cargo install` wrote to ~/.cargo/bin. On this runner ~/.local/bin
           # comes FIRST in PATH and held a 24-day-old apr, so every previous run
           # of this story installed a fresh binary and then executed a stale one:
           #
@@ -96,8 +106,8 @@ jobs:
           # Detected because the format_parity gate reported its OLD result
           # shape (value/threshold 17/17, a token id) on a commit whose source
           # reports 64/64 decode steps. Put the freshly built binary first.
-          echo "$HOME/.cargo/bin" >> "$GITHUB_PATH"
-          export PATH="$HOME/.cargo/bin:$PATH"
+          echo "$APR_BUILT_DIR" >> "$GITHUB_PATH"
+          export PATH="$APR_BUILT_DIR:$PATH"
 
           # Fail closed on staleness. A path fix alone would silently rot again
           # the next time PATH changes; this asserts that the apr the story is

.quorum/0694-batch1/receipt-r2-c13a233a3.json

@@ -0,0 +1,63 @@
+{
+  "pr": 4318,
+  "batch": "batch/0.69.4-epic-4249",
+  "reviewed_head": "c13a233a3",
+  "base": "3c4e19242 (last quorumed, .quorum/0694-batch1/receipt.json)",
+  "author_model": "claude-opus-5-5 (fold captain aprender-59); fix author aprender-0d",
+  "degraded": false,
+  "round_1": {
+    "head": "d1bb4d9f4",
+    "range": "3c4e19242..d1bb4d9f4",
+    "diff_sha256": "8f79bb1c8f91446a250a9f3c42ece2f051f0698377b2c1790afc536e8b2843f3",
+    "lanes": [
+      {
+        "lane": 1,
+        "model": "claude-sonnet-5",
+        "verdict": "PASS"
+      },
+      {
+        "lane": 2,
+        "model": "claude-haiku-4-5-20251001",
+        "verdict": "PASS"
+      },
+      {
+        "lane": 3,
+        "model": "gemini-3.1-pro-high",
+        "model_self_report": "Gemini 3.1 Pro",
+        "verdict": "FAIL",
+        "finding": "qwen35_session_tests.rs:575/606 test grown_capacity and allocate_with_exact_retry only; ensure_capacity (qwen35_session.rs ~620-668), their caller, is driven by no test, so reverting its call site stays green"
+      }
+    ]
+  },
+  "fix": {
+    "commit": "c13a233a3",
+    "author": "aprender-0d",
+    "what": "CPU test drives ensure_capacity through reserve(); GPU test with a cfg(test) alloc ceiling + control; mutant reverting the grown_capacity call site is RED (20 vs 32)"
+  },
+  "round_2_on_fix": {
+    "head": "c13a233a3",
+    "range": "d1bb4d9f4..c13a233a3",
+    "diff_sha256": "ad23badb311c012c14302379da3d432fcd0b6fa7954edc0e8ff7fef691c52761",
+    "lanes": [
+      {
+        "lane": 1,
+        "model": "claude-sonnet-5",
+        "verdict": "PASS"
+      },
+      {
+        "lane": 2,
+        "model": "claude-haiku-4-5-20251001",
+        "verdict": "PASS"
+      },
+      {
+        "lane": 3,
+        "model": "gemini-3.1-pro-high",
+        "model_self_report": "Gemini 3.1 Pro (High)",
+        "verdict": "PASS",
+        "agy_status": "SUCCESS",
+        "agy_conversation": "7bba17fe-c966-445c-ad5b-c8be7b969847"
+      }
+    ]
+  },
+  "result": "PASS 3/3 at c13a233a3 (round 1 FAIL resolved by c13a233a3)"
+}

.quorum/0694-batch1/receipt.json

@@ -0,0 +1,23 @@
+{
+  "pr": 4318,
+  "batch": "batch/0.69.4-epic-4249",
+  "reviewed_head": "01c71e648",
+  "fix_head": "3c4e19242",
+  "base": "a5efa1443",
+  "author_model": "claude-opus-5-5",
+  "degraded": "same-family",
+  "mandatory_scope": ".github/workflows/binary-release.yml reviewed explicitly by every lane (cop GO condition)",
+  "round_1": [
+    {"lane": 1, "model": "claude-sonnet-5", "scope": "workflows + scripts/release + install", "verdict": "FAIL", "finding": "build-apr-darwin did not stamp the rc version; the rc darwin asset printed bare X.Y.Z"},
+    {"lane": 2, "model": "claude-haiku-4-5-20251001", "scope": "binary-release.yml line-by-line + ci/ + contracts/", "verdict": "PASS"},
+    {"lane": 3, "model": "claude-sonnet-5", "scope": "crates/** + binary-release.yml", "verdict": "PASS"}
+  ],
+  "fix": "3c4e19242: darwin lane stamps the rc version (stamper from the workflow commit) and runs asset_version_check.sh against the tag and commit",
+  "round_2_on_fix": [
+    {"lane": "r1", "model": "claude-sonnet-5", "verdict": "FAIL", "reason": "darwin is not a required asset and has no post-upload smoke; this is aprender#4353 scope (milestone 0.70.1), and the operator ruled no scope change for 0.69.4"},
+    {"lane": "r2", "model": "claude-haiku-4-5-20251001", "verdict": "PASS"},
+    {"lane": "r3", "model": "claude-sonnet-5", "verdict": "PASS", "note": "#4353 gap is non-blocking and out of scope; smoke the stamper on the mini"}
+  ],
+  "extra_evidence": "stamp_rc_version.sh --self-test on mini (GNU bash 5.3.20, aarch64-apple-darwin): PASS",
+  "result": "3/3 PASS (lanes 2 and 3 on the whole batch; r2 and r3 on the fix to lane 1's finding), degraded: same-family"
+}