spec(gemma4): repair #140 integration regressions

[localai-bot]

Scope

Committed written repair plan only for MODEL-TEXT-gemma4-gemma4-for-causal-lm. No production or test fix is included.

Spec: .agents/specs/main-regression-repair-2026-08-08.md
Immutable reviewed head: e0b185c

Corrected root causes

1. gemma4_moe.cpp has two direct vt::GeluAndMul calls, not SiluAndMul. They violate the shared merged-GEMM/MoE MlpGateUpMethodBase family; this is not the residual-norm FusedChain glue seam.
2. runner.cpp hard-codes kCUDA in shared async-input eligibility despite the device-leakage baseline binding kcuda=0. Eligibility must be unified host readability OR a Platform-advertised device mirror.

The plan forbids an allowlist, DSR waiver, baseline increase, or #140 revert; requires two independent TDD commits and mutation review; then merges the repair before reconciling #141 keyed records wholesale from current main.

Intake

• feat(rocm): gfx1201 hipBLAS ops + Gemma-4-26B-A4B MoE (BF16/FP8) #140 is merged and introduced both affected files.
• perf(gemma4): dual-GPU FP8 resident without host OOM + peer mix #154 overlaps gemma4_moe.cpp but is a separate resident-expert performance campaign; implementation must re-inspect and reconcile its current head.
• No open issue or row PR already claimed this repair at intake.

Validation

Green: doc-checkpoint staged/commit/range, public-doc checker + 41 mutations, agent-record + 13 mutations, protocol consistency, NOW-current, diff check, PR-size (0 non-exempt lines).

Expected red-first baseline at assigned base 1ce0d66: check-fusion-consistency and test_check_fusion_consistency both name gemma4_moe. Current main later added an allowlist entry; the plan requires removing it as part of the real repair.

Draft remains stopped for written-spec review.

FOLLOWING_AGENTS_PROTOCOL