ci: drop X64 from arch-neutral runs-on, pin + assert x86_64 on the FULL tier (#5871 re-cut, #5882 blocker)

.github/workflows/ci.yml

@@ -102,7 +102,7 @@ jobs:
     # its own. It runs on push only; on merge_group it is skipped at t=0, which the verdict blocks
     # below read as `skipped:::skipped` = "x86-main + determinism ran".
     if: github.event_name == 'push' && github.ref == 'refs/heads/main'
-    runs-on: [self-hosted, Linux, X64, clean-room]
+    runs-on: [self-hosted, Linux, clean-room]
     timeout-minutes: 10
     permissions:
       contents: read
@@ -316,7 +316,7 @@ jobs:
   x86-main-advisories:
     needs: [mg-reuse]
     if: ${{ !cancelled() && needs.mg-reuse.outputs.x86 == '1' }}
-    runs-on: [self-hosted, Linux, X64, clean-room]
+    runs-on: [self-hosted, Linux, clean-room]
     timeout-minutes: 30
     permissions:
       contents: read
@@ -439,12 +439,35 @@ jobs:
   # failure, never a pass.
   # ==========================================================================
   # The workspace tests, one shard per job (nextest hash partition N/12 plus the
-  # explicit runner's round-robin N/12), in parallel on the X64 clean-room pool.
+  # explicit runner's round-robin N/12), in parallel on the clean-room pool (arch-neutral, PMAT-3138).
   # The section body is ci/sections.yml `workspace-test-shard`, unchanged.
+  # #5882: which tier this run owes decides WHERE the shards may run. The full tier's Σ-executed fan-in
+  # compares every shard with ONE universe list, and arch-gated tests (target_arch cfgs) differ per
+  # arch, so a full tier needs one architecture: X64. The quick/PR tier owes no Σ and stays
+  # arch-neutral. The tier is only knowable from the diff, so it is decided HERE, by the same script
+  # the shards run (scripts/ci/decide_tier.sh), and workspace-test-shard reads it in its runs-on.
+  # scripts/ci/sigma_arch_uniform.sh in the fan-in is the backstop: a mixed full tier is RED.
+  workspace-test-arch:
+    runs-on: [self-hosted, Linux, clean-room]
+    timeout-minutes: 15
+    outputs:
+      tier: ${{ steps.tier.outputs.tier }}
+    steps:
+      - uses: actions/checkout@v7
+        with:
+          fetch-depth: 0
+      - name: Decide the test tier (the shards' own step, one script)
+        id: tier
+        env:
+          GH_TOKEN: ${{ github.token }}
+          MG_BASE_SHA: ${{ github.event.merge_group.base_sha }}
+        run: bash scripts/ci/decide_tier.sh
   workspace-test-shard:
+    needs: [workspace-test-arch]
     if: ${{ github.event.pull_request.draft != true }}
-    # same lanes as x86-main: clean-room for pull_request/push, mq for a merge_group entry
-    runs-on: ${{ github.event_name == 'merge_group' && fromJSON('["self-hosted","Linux","X64","mq"]') || fromJSON('["self-hosted","Linux","X64","clean-room"]') }}
+    # arch-neutral (operator 2026-10-21 14:18Z) EXCEPT the full tier, which needs one arch for the Σ (#5882, workspace-test-arch)
+    # clean-room for pull_request/push, mq for a merge_group entry
+    runs-on: ${{ needs.workspace-test-arch.outputs.tier == 'full' && (github.event_name == 'merge_group' && fromJSON('["self-hosted","Linux","X64","mq"]') || fromJSON('["self-hosted","Linux","X64","clean-room"]')) || (github.event_name == 'merge_group' && fromJSON('["self-hosted","Linux","mq"]') || fromJSON('["self-hosted","Linux","clean-room"]')) }}
     timeout-minutes: 180
     strategy:
       fail-fast: false
@@ -523,6 +546,7 @@ jobs:
           set -euo pipefail
           d="$RUNNER_TEMP/sigma"
           python3 scripts/ci/sigma_executed.py --self-test
+          bash scripts/ci/sigma_arch_uniform.sh --self-test
           # ONE shard count: the pins in ci/sections.yml (check_shard_count_single_source.sh)
           SHARDS="$(awk -v ev="$GITHUB_EVENT_NAME" '/^matrix-pins:/{p=1;next} p&&/^  workspace-test-shard:/{k="pr";next} p&&/^  workspace-test-shard-merge_group:/{k="mg";next} p&&/^  [a-z]/{k=""} k&&/^    - \{shard:/{c[k]++} END{print (ev=="merge_group"&&c["mg"]?c["mg"]:c["pr"])+0}' ci/sections.yml)"
           [ "${SHARDS:-0}" -ge 1 ] 2>/dev/null || { echo "::error::Σ: no workspace-test-shard pins"; exit 1; }
@@ -539,6 +563,8 @@ jobs:
           [ "$#" -eq "$SHARDS" ] || { echo "::error::Σ: $# tier files for $SHARDS shards:$tiers"; exit 1; }
           for t in "$@"; do [ "$t" = "$1" ] || { echo "::error::Σ: the shards disagree on the tier:$tiers"; exit 1; }; done
           [ "$1" = full ] || { echo "every shard ran tier '$1': the lib/explicit Σ is not owed"; exit 0; }
+          # #5882: a full tier on two architectures is RED (arch-gated tests, one universe): placement is workspace-test-arch's, this is the backstop
+          bash scripts/ci/sigma_arch_uniform.sh "$d" "$SHARDS"
           j=() l=()
           for n in $(seq 1 "$SHARDS"); do
             j+=(--junit "$d/sigma-shard-$n/lib.junit.xml")

ci/sections.yml

@@ -378,66 +378,7 @@ jobs:
           GH_TOKEN: ${{ github.token }}
           MG_BASE_SHA: ${{ github.event.merge_group.base_sha }}
         run: |
-          set -euo pipefail
-          args=(--event "${GITHUB_EVENT_NAME}")
-          # PMAT-1098 67-E2: the queue and a push to main re-derive their own
-          # diff from HEAD's FIRST PARENT (HEAD^1..HEAD is the PR's diff on the
-          # new base). actions/checkout is shallow, so that parent is not in the
-          # object store until we deepen -- and without it ci_test_tier.sh falls
-          # closed to `full` every single time, which would make the whole
-          # optimisation inert while looking like it worked. Best effort by
-          # design: if the deepen fails the decision still falls closed, and the
-          # warning below says so instead of the run silently paying the hour.
-          case "${GITHUB_EVENT_NAME}" in
-            merge_group)
-              # Operator 2026-10-17: the queue runs the quick tier over the entry's diff (base_sha = prior entry's head),
-              # MG_BASE_SHA..HEAD. Fetch the base; if that fails the tier falls closed to full.
-              bash scripts/ci/fetch_p6.sh --no-tags --depth=1 origin "${GITHUB_SHA}" "${MG_BASE_SHA}" 2>/dev/null \
-                || bash scripts/ci/fetch_p6.sh --no-tags --depth=1 origin "${MG_BASE_SHA}" 2>/dev/null \
-                || true
-              git cat-file -e "${MG_BASE_SHA}^{commit}" 2>/dev/null \
-                || printf '::warning::merge_group base sha unavailable -- the tier falls closed to full\n'
-              ;;
-            push)
-              bash scripts/ci/fetch_p6.sh --no-tags --depth=2 origin "${GITHUB_SHA}" 2>/dev/null \
-                || bash scripts/ci/fetch_p6.sh --no-tags --deepen=1 origin 2>/dev/null \
-                || true
-              if git rev-parse -q --verify 'HEAD^1' >/dev/null 2>&1; then
-                printf 'first parent available: %s\n' "$(git rev-parse HEAD^1)"
-              else
-                printf '::warning::first parent unavailable after deepen -- the tier falls closed to full\n'
-              fi
-              ;;
-          esac
-          case "${GITHUB_EVENT_NAME}" in
-            pull_request)
-              bash scripts/ci/fetch_p6.sh --no-tags --depth=1 origin "+refs/heads/${GITHUB_BASE_REF}:refs/remotes/origin/${GITHUB_BASE_REF}"
-              # #5391: the PR's change is the diff from merge-base(base, HEAD), not the
-              # base tip -- main's later commits are not the PR's. --no-renames (#3664)
-              # lives in the helper.
-              bash scripts/ci/pr_touched_paths.sh "origin/${GITHUB_BASE_REF}" HEAD > "$RUNNER_TEMP/touched.txt"
-              args+=(--diff-from "$RUNNER_TEMP/touched.txt")
-              ;;
-            merge_group)
-              # Operator 2026-10-17 (cop window), verbatim: "only gating on tags/release".
-              # ci_test_tier.sh diffs MG_BASE_SHA..HEAD (env above) and selects like a PR.
-              ;;
-            push)
-              # T36: the merge queue already tested this very sha. Hand its
-              # workspace-test result to the tier script, which reuses it only
-              # on a success for HEAD itself. A failed lookup is the push's own
-              # tier, as before -- never a reuse, never a red step.
-              mg=$(bash scripts/ci_mg_workspace_result.sh "${GITHUB_REPOSITORY}" "${GITHUB_SHA}") || mg=""
-              if [ -n "$mg" ]; then
-                mapfile -t mgargs <<<"$mg"
-                args+=("${mgargs[@]}")
-              else
-                printf 'no merge_group workspace-test result for %s: the tier is decided from the push diff\n' "${GITHUB_SHA}"
-              fi
-              ;;
-          esac
-          bash scripts/ci_test_tier.sh "${args[@]}" | tee "$RUNNER_TEMP/tier.txt"
-          grep -E '^(tier|crates|targets|check_workspace|reason|cite)=' "$RUNNER_TEMP/tier.txt" >> "$GITHUB_OUTPUT"
+          bash scripts/ci/decide_tier.sh
       # #4502 row 2: lint_passes_on_real_contracts (aprender-contracts) requires the
       # refinement gate to RUN whenever CI is set, and that gate compares HEAD with
       # merge-base(HEAD, origin/main), else the origin/main tip. This section clones at
@@ -774,6 +715,7 @@ jobs:
           sig="$RUNNER_TEMP/sigma"
           mkdir -p sigma-shard
           printf '%s\n' "$TIER" > sigma-shard/tier
+          uname -m > sigma-shard/arch
           if [ "$TIER" = full ]; then
             cp "$sig/lib.junit.xml" sigma-shard/lib.junit.xml
             cp "${FAT_SECTION_LOG:?the fat driver sets FAT_SECTION_LOG}" sigma-shard/section.log

crates/aprender-serve/src/kernel_registry_parity.rs

@@ -743,6 +743,7 @@ fn committed() -> Vec<(serde_json::Value, serde_json::Value)> {
 #[test]
 fn committed_parity_receipts_hold_on_this_host() {
     let r = registry().expect("registry");
+    let mut no_arch_receipt = Vec::new();
     for (entry, rc) in committed() {
         let id = rc["kernel_id"].as_str().expect("kernel_id");
         let path = entry["receipt"].as_str().expect("receipt path");
@@ -789,8 +790,22 @@ fn committed_parity_receipts_hold_on_this_host() {
             seed: wl["seed"].as_u64().expect("seed"),
             trials: dim("trials"),
         };
-        let bound = rc["tolerance_rel"].as_f64().expect("tolerance_rel");
         let now = measure(kernel(id), w, false);
+        // A receipt's `tolerance_rel` is the bound measured on its own `host_arch`; libm and FMA
+        // contraction move the last ulps between architectures. Any other arch must carry its own
+        // measured bound under `per_arch.<arch>.tolerance_rel` -- never a borrowed one, never a skip.
+        let bound = if rc["host_arch"] == host_arch() {
+            rc["tolerance_rel"].as_f64().expect("tolerance_rel")
+        } else if let Some(b) = rc["per_arch"][host_arch()]["tolerance_rel"].as_f64() {
+            b
+        } else {
+            no_arch_receipt.push(format!(
+                "{path}: no {} receipt (host measured max_rel_err {})",
+                host_arch(),
+                now.max_rel_err
+            ));
+            continue;
+        };
         assert_eq!(
             rc["margin"]["verdict"], "pass",
             "{path}: KTEST-02 margin verdict (max ≤ 1 under EM-DOT)"
@@ -842,6 +857,11 @@ fn committed_parity_receipts_hold_on_this_host() {
             p => panic!("{id}: no precision check for {p}"),
         }
     }
+    assert!(
+        no_arch_receipt.is_empty(),
+        "committed receipts lack a bound for this arch:\n{}",
+        no_arch_receipt.join("\n")
+    );
 }
 
 /// `KREG_INPUT_SETS_OUT`: the input-set hashes recomputed from the tree at `KREG_GIT_SHA`, one

evidence/kreg/parity/cpu.matvec.apr_q4.json

@@ -35,6 +35,15 @@
   },
   "oracle": "apr_format_decode_f64",
   "oracle_independent": true,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 5.568608963671898e-07,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 2e-6,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": null,
   "registry_precision": "f32",
   "schema": "kernel-parity-receipt/v1",

evidence/kreg/parity/cpu.matvec.apr_q8.json

@@ -35,6 +35,15 @@
   },
   "oracle": "apr_format_decode_f64",
   "oracle_independent": true,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 5.982009671170112e-07,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 2e-6,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": null,
   "registry_precision": "f32",
   "schema": "kernel-parity-receipt/v1",

evidence/kreg/parity/cpu.matvec.bf16.json

@@ -33,6 +33,15 @@
   },
   "oracle": "le_decode_half_f64",
   "oracle_independent": true,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 4.068146011742945e-07,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 9e-7,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": null,
   "registry_precision": "f32",
   "schema": "kernel-parity-receipt/v1",

evidence/kreg/parity/cpu.matvec.f16.json

@@ -33,6 +33,15 @@
   },
   "oracle": "le_decode_half_f64",
   "oracle_independent": true,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 3.920968817889049e-07,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 8e-7,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": null,
   "registry_precision": "f32",
   "schema": "kernel-parity-receipt/v1",

evidence/kreg/parity/cpu.matvec.f32.json

@@ -33,6 +33,15 @@
   },
   "oracle": "le_decode_half_f64",
   "oracle_independent": true,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 7.356022484897627e-07,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 2e-6,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": null,
   "registry_precision": "f32",
   "schema": "kernel-parity-receipt/v1",

evidence/kreg/parity/cpu.matvec.q4_0.json

@@ -35,6 +35,15 @@
   },
   "oracle": "in_tree_dequant_f64",
   "oracle_independent": false,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 0.003859434297246209,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 0.008,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": {
     "max_rel_err": 1.69613194525447e-7
   },

evidence/kreg/parity/cpu.matvec.q4_k.json

@@ -36,6 +36,15 @@
   },
   "oracle": "in_tree_dequant_f64",
   "oracle_independent": false,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 0.005569836531516333,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 0.02,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": {
     "max_rel_err": 1.3720446505396093e-7
   },

evidence/kreg/parity/cpu.matvec.q5_k.json

@@ -36,6 +36,15 @@
   },
   "oracle": "in_tree_dequant_f64",
   "oracle_independent": false,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 1.3110594897372703e-06,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 3e-6,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": null,
   "registry_precision": "f32",
   "schema": "kernel-parity-receipt/v1",

evidence/kreg/parity/cpu.matvec.q6_k.json

@@ -36,6 +36,15 @@
   },
   "oracle": "in_tree_dequant_f64",
   "oracle_independent": false,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 1.6505971099429934e-6,
+      "source": "CI job 119280981403 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 4e-6,
+      "ts": "2026-10-21T14:37:43Z"
+    }
+  },
   "quantized_activation_oracle": null,
   "registry_precision": "f32",
   "schema": "kernel-parity-receipt/v1",

evidence/kreg/parity/cpu.matvec.q8_0.json

@@ -36,6 +36,15 @@
   },
   "oracle": "in_tree_dequant_f64",
   "oracle_independent": false,
+  "per_arch": {
+    "aarch64": {
+      "host": "gx10-pool1",
+      "max_rel_err": 0.003738839019401381,
+      "source": "CI job 119289824035 (PR #5871 shard 1), same seed and workload",
+      "tolerance_rel": 0.008,
+      "ts": "2026-10-21T14:51:19Z"
+    }
+  },
   "quantized_activation_oracle": {
     "max_rel_err": 2.0240994602880252e-7
   },

scripts/check_ci_push_reuse.sh

@@ -285,10 +285,13 @@ hasnt "D11 pull_request ignores --mg-*: never reuse" "tier=reuse" "$OUT"
 
 # -- W: the CI wiring ------------------------------------------------------------------------------
 SEC="$ROOT/ci/sections.yml"
+DT="$ROOT/scripts/ci/decide_tier.sh" # the tier step body lives in one script (#5882): the section step and workspace-test-arch both run it
+want "W0 the section step runs the shared decide_tier.sh" 1 \
+     "$(grep -cF "bash scripts/ci/decide_tier.sh" "$SEC")"
 want "W1 the tier step's push arm calls the lookup for GITHUB_SHA" 1 \
-     "$(grep -cF 'bash scripts/ci_mg_workspace_result.sh "${GITHUB_REPOSITORY}" "${GITHUB_SHA}"' "$SEC")"
+     "$(grep -cF 'bash scripts/ci_mg_workspace_result.sh "${GITHUB_REPOSITORY}" "${GITHUB_SHA}"' "$DT")"
 want "W2 a failed lookup is not fatal to the step (it falls to today's tier)" 1 \
-     "$(grep -cF '"${GITHUB_SHA}") || mg=""' "$SEC")"
+     "$(grep -cF '"${GITHUB_SHA}") || mg=""' "$DT")"
 W3PAT="push) printf 'cited merge_group run: %s"
 if grep -qF "$W3PAT" "$SEC"; then w3=1; else w3=0; fi
 want "W3 the reuse step labels a push citation as a merge_group run, not a PR head" 1 "$w3"