[noahgift]

One commit on current main (0a100ed) combining 8383de3 and 748da4069c. Supersedes #5886.

• Arch-neutral runs-on loses X64, so ARM64 clean-room runners can take PR shards. workspace-test-arch decides the tier first and pins the full tier's shards to X64.
• scripts/ci/sigma_arch_uniform.sh: at tier full every shard must have staged x86_64; a missing or blank arch, or a blank tier, is RED. --self-test: 12/12 rows as expected (local run).
• Kernel parity receipts carry per_arch.aarch64 bounds measured on gx10-pool1 (CI job 119289824035) instead of borrowing the x86 bound; a host with no receipt for its arch fails, never skips.

Order (operator 17:46Z, relayed by aprender-57): merge queue right after #5883, ahead of #5885. Quorum run by triage.

🤖 Generated with Claude Code

[noahgift]

seat opus-5-5-lane-a APPROVE @0a100ed865

[noahgift]

quorum 3/3 AGREED PASS @0a100ed865

[noahgift]

seat opus-5-5 APPROVE @0a100ed865

• Head verified: gh pr view 5891 headRefOid = 0a100ed, 1 commit, 17 files.
• Differs from #5882 on the known blocker: #5882 only dropped X64. #5891 adds job workspace-test-arch (runs the shared scripts/ci/decide_tier.sh), and workspace-test-shard pins X64 when tier==full (mq or clean-room), arch-neutral otherwise. Each shard stages uname -m, and the fan-in runs sigma_arch_uniform.sh after the tier-agreement check: at full tier, any non-x86_64, blank or missing arch, or a blank tier, is RED. The triage receipt (triage-quorum-5891-0a100ed865.md, 3/3 PASS) covers this.
• Fails closed: if the arch job fails, the shards are skipped, and workspace-test treats skipped as not a pass (ci.yml:522-525). If the tier drifts (arch job says quick, shards say full), the backstop turns RED. No gate weakened: the x86 tolerance_rel values are unchanged. On another arch, kernel_registry_parity now asserts on a missing per_arch bound (never a skip). The aarch64 bounds are ≥ their own measured max_rel_err (q4_k 0.02 vs 0.0056, f32 2e-6 vs 7.4e-7, bf16 9e-7 vs 4.1e-7).
• Executed on scratchpad copies of the head (git show): sigma_arch_uniform.sh --self-test 12/12, including the mixed placement from run 39620701544 and all-aarch64 → rc1. ci_mg_reuse.sh --self-test 108/0 (adds 3 pool mutants: full-tier-loses-x64-pin, every-tier-pinned-x64, tier-job-off-the-pool). check_ci_push_reuse.sh rc0.
• Scope: the quick/PR/MQ tier loses x86-only tests. That follows the operator's 14:18Z arch-neutral ruling and is not a defect. The step body moved verbatim from sections.yml into decide_tier.sh.
• Non-blocking, agreeing with the receipt: the ci_test_tier.sh --self-test row for pr_touched_paths still points at sections.yml (LAB red in guards-nightly; repoint it to decide_tier.sh). sigma_arch_uniform takes SHARDS<1 as green (the caller already rejects it). kernel_registry_op_parity.rs still applies the x86 bound on aarch64 (needs a ticket). The extra serial arch job's cost was not measured.

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=5891 head=815d4e7f1db4f659f0e226b4a65ddb5b824fc0d6 verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.

[noahgift]

seat opus-5-5-lane-a APPROVE @815d4e7f1d

[noahgift]

quorum 3/3 AGREED PASS @815d4e7f1d

[noahgift]

seat opus-5-5 APPROVE @815d4e7f1d

• Head: gh pr view 5891 headRefOid = 815d4e7, OPEN, MERGEABLE. compare 0a100ed...815d4e7 = ahead 2, behind 0: main's f0d48dc (draft-skip) plus the merge commit 815d4e7 (parents 0a100ed, f0d48dc). Nothing was force-pushed.
• Design intact in gh pr diff 5891: the workspace-test-arch job runs scripts/ci/decide_tier.sh and outputs tier. workspace-test-shard has needs [workspace-test-arch], and its runs-on pins X64 (mq or clean-room) only when tier == 'full', otherwise it stays arch-neutral. The fan-in runs sigma_arch_uniform.sh --self-test and then the backstop on "$d" "$SHARDS". The ci_mg_reuse.sh pool mutants (full-tier-loses-x64-pin, every-tier-pinned-x64, tier-job-off-the-pool) are all still present.
• Merge composition at the head's ci.yml: the shard keeps needs [workspace-test-arch] and also gets main's draft if:. The fan-in workspace-test (needs shard, if always()) exits 0 only when DRAFT == true. Otherwise it is RED unless SHARD_RESULT == success, so a skipped shard, or a failed arch job that skips the shards, is still RED. The Σ checks for tier-file presence and count, tier agreement and arch uniformity are unchanged. No gate row was dropped or weakened.
• Minor, not blocking: workspace-test-arch has no draft if:, so it still runs a short tier decision (15-min timeout) on drafts. It only costs runner time and changes no gate.
• Triage receipt triage-quorum-5891-815d4e7f1d.md agrees: the delta is only the upstream draft-skip, and on the head it records sigma_arch_uniform self-test 12/12, ci_mg_reuse 108/0, and push_reuse, guards_wired and ci_draft_skip all rc 0. I checked its claims against the diff; I did not run the self-tests again here.