fix(seedless): check command buffer status, poison the backend on GPU failure

[penta2himajin]

Refs #169. Stops the silent failure; does not fix the OOM itself.

What this changes

A failed Metal command buffer leaves its output buffers untouched. Shared buffers are zero-initialised, so reading them anyway yields an all-zero logit row, which argmax_rows tie-breaks to index 0 — token id 0 (!), forever. Nothing inspected cb.status before this: 94 waitUntilCompleted call sites, zero error checks.

• commitAndWaitChecked(cb, label) -> String? — nil on success, a fault description otherwise. Non-nil means the caller must not read that command buffer's outputs.
• GPUFaultState — per-instance, sticky, first-fault-wins. Deliberately not process-global: one failed request must not kill a long-running server. SeedlessFusedForward owns one.
• Applied to the decode readback funnels — stepArgmax (.resident/.bolt) and chainedStepArgmax — plus an entry guard so a poisoned backend refuses rather than reading stale buffers. The chain advances CPU-side KV length at encode time, so continuing after a failure is unsound, not merely wasteful.
• Decode path selection is now logged unconditionally. The same logical op has three implementations (raw fused step, GPU chain, MLX-composed fallback) and which one runs is a runtime decision. Without this line, instrumentation lands on a path that never executes — that cost most of the debugging time on bolt: decode command buffer fails with GPU OOM and cb.error is never checked → unbounded token-id-0 (!) output above ~2–6K prompt tokens #169.
• QWISP_LOGIT_DBG (default off): per-row logit/hidden stats and cb.status.

Evidence

16GB-class config, one prose prompt truncated to a ladder, cb.error read at the failing commit:

config	fails at
bolt mixed C=128	~4.0–4.3K prompt tok
bolt generic C=128	between 6.6K and 14.7K
strict (8GB budget, C=64)	passes 41,489 tok

At every failure: MTLCommandBufferErrorDomain Code=8 kIOGPUCommandBufferCallbackErrorOutOfMemory, logits all zero (nan=0), while hBuf and normed held healthy leftover values from prefill — the CB simply never ran.

Not in this PR, deliberately

• strict's split command buffers, stepArgmaxBatch, prefill CBs
• a static gate against bare waitUntilCompleted
• an error channel on LLMBackend.generate's AsyncStream — today the failure stops generation and reports on stderr, but does not reach the caller as a typed error, so the OpenAI-compatible surface cannot yet return a proper error response
• the capacity bug itself (why the CB does not fit) — stays open in bolt: decode command buffer fails with GPU OOM and cb.error is never checked → unbounded token-id-0 (!) output above ~2–6K prompt tokens #169

Gates

RAWTESTS 99/99. The new locked test asserts the checked-commit contract and that the poison is per-instance, sticky, and does not leak between backends. Known limitation: failure is injected via a test flag, so removing only the cb.status comparison while keeping the injection check would not be caught — the static gate in the follow-up should close that.

🤖 Generated with Claude Code