fix(serve): multi-slot engine fits 24 GB cards — skip GDN S-tape, chunk prefill scratch, apply device visibility

[alpineQ]

Summary

Fixes #602. Three fixes that together let serve.multi_slot start on a consumer 24 GB card:

1. Skip the GDN S-tape in the engine's prefill scratch. Rig::build called PrefillBatchScratch::new(), which allocates the DDTree tree-verify S-tape sized [max_batch × n_v_heads × 128²] (f32 + q8) — ~21 GB at 16k ctx. The slots path is plain prefill and never tree-verifies; new_opt's own docstring prescribes cap_gdn_tape=false for exactly this case.
2. Size scratch by n_slots × prefill_chunk, not cap_tokens. The scheduler was built around chunked prefill (its docstring, and the SP3 spec's "slot 1 is chunk-prefilling 256 tokens" example), but run_loop set chunk_size = cap_tokens, which both forced cap-sized scratch and let one long prompt monopolise a step. EngineConfig.prefill_chunk (from serve.multi_slot_prefill_chunk, default 1024) now bounds both.
3. Apply device visibility in serve. The engine runs in-process and never executed the daemon's apply_device_visibility, so hardware.devices / HIPFIRE_DEVICES were ignored and the engine always took device 0 — the display card on a desktop.

Result: qwen3.6-35b-a3b.mq4r at 2 slots × 16384 ctx runs in 23.4 GB on a 24 GB gfx1100 (any usable ctx OOM'd before), pinned to the configured device.

Stacked on #607 (which registers serve.multi_slot_prefill_chunk) — review that first; this branch contains its commit.

Which crate(s) does this touch?

• crates/hipfire-arch-qwen35
• examples / daemon (hipfire-cli serve, bench_concurrency, test_serve_concurrent — EngineConfig constructors)

Test plan

• cargo build --release --workspace --features deltanet clean
• cargo test --lib --workspace --features deltanet passes
• GPU gates on gfx1100 24 GB with qwen3.6-35b-a3b.mq4r: test_prefix_cache_equivalence, test_swap_roundtrip, test_swap_equivalence, test_serve_concurrent — all pass (none of them can even start on beta at this ctx: the engine OOMs first)
• attn_legacy_baseline.sh bitwise-identical to beta (single-sequence path untouched)
• Paired fresh-process decode bench beta-vs-branch (qwen3.5-4b.mq4, asym3 + graph, 1 warm + 2 measured per side): 175.7 / 175.6 vs 174.7 / 174.7 tok/s → −0.55%, inside the ±1–3% session band
• speed.arch-fast passes in the gate below
• Change gate run and telemetry pasted below

change_gate telemetry

**change_gate: FAIL**

host gfx=`gfx1100` rocm=`6.18.37` models_dir=`/home/alpineq/.hipfire/models` · `80a572c824dcbbcdddac1cc5f6d1a7e91d7c4dd6`..`0d50f8b1a7cb41fec550659dc53a76baa6f46ff3` dirty · est=14.05min actual=37.0s

### Routes RUN
| route                   | status | duration |
| ----------------------- | ------ | -------- |
| serve.battery.qwen35-4b | fail   | 0.1s     |
| serve.battery.qwen35-9b | fail   | 0.1s     |
| speed.arch-fast         | fail   | 19.5s    |
| unit.arch-qwen35        | pass   | 1.7s     |
| unit.diff-check         | pass   | 0.0s     |
| unit.env-docs           | fail   | 0.2s     |
| unit.hipfire-cli        | pass   | 1.9s     |
| unit.hipfire-config     | pass   | 0.1s     |
| unit.hipfire-runtime    | pass   | 10.8s    |
| unit.no-gpu-control     | fail   | 2.5s     |

### Routes NOT RUN
| route | reason |
| ----- | ------ |
| —     | —      |

_Blocked or excluded routes mean coverage is incomplete — this report is not an admission that unrun surfaces are safe._

Pre-existing failures on this host, verified against a clean `upstream/beta` worktree and NOT introduced here:

• unit.env-docs — 173 findings, every one in a crate this diff does not touch (quantize, gemma4, deepseek4, redline, ...). No finding names a file in this PR.
• unit.no-gpu-control — hipfire-tui's bounded_run_stdout_fill_is_not_falsely_timed_out fails identically on pristine beta (timing-sensitive dashboard test).
• serve.agentic.a3b-fast — zero tokens emitted on pristine beta too.
• serve.battery.qwen35-4b / -9b — the route argv asks --thinking med with max_tokens 180/300, which serve_harness.py rejects before touching the GPU (the harness defect PR fix(serve): emit Qwen reasoning content #572 documented).