#2224 pcie: the measured PCIe share default (PcieAuto) on gfx1100 too — gate-measured on a native pack

Description

0.1.43 made PcieAuto the default under the gfx12 decode defaults on HIP (`decode_defaults_arch_ok` → gfx1200/gfx1201, measured on 2×R9700) and for the canonical pack on CUDA (ab97f174). Every other AMD part keeps its earlier behaviour.

This extends the default to **gfx1100 (RX 7900 XT/XTX/GRE)** — the one other arch where the in-run refinement has been gate-measured, on a native (IQ) pack:

**No-regression gate** (4 alternating engine starts @64K, same args/env, paired medians): prefill 2189 (base) vs 2180 (with the arm), decode 65.4 vs 65.4 — inside noise. The gate's rule was no-regression, so the arm stays.

**Engagement** (under `--serve`, real ~95K-token request): the in-run refinement fires and prints `pool 42-45 us/missed expert (45-48 GB/s of expert bytes), the link 28.4-28.5 GB/s -> pcie_frac 0.25 (was 0.55)`, self-limited to 3 updates — the engine's own `f = L/(L + 1.85 r)`. It has run as our daily-driver production config since 2026-10-12 and was re-verified after the 0.1.43.1 refresh.

Scope, deliberately narrow:
- **gfx1100 only** — `pcie_auto_fork_arch_ok()` pins the arch string exactly. gfx1200/1201 keep 0.1.43's default; every other AMD part keeps the gate's earlier behaviour. Nothing is claimed for unmeasured archs.
- **Native packs only** — the canonical pack keeps its CUDA-only refinement, including the Windows opt-in (STRATA_PCIE_AUTO_CANON).
- Opt-outs unchanged: `STRATA_PCIE_FRAC_DEFAULT=old`, `--pcie-frac N`, the layer-split rules (incl. the `STRATA_PCIE_AUTO_SPLIT` opt-in).

Build: HIP gfx1100, 373/373 targets on the nightly ROCm 10.2 toolchain. The arm is host-side policy (no kernel change); the CUDA path is untouched — `pcie_auto_fork_arch_ok()` returns false there and `decode_defaults_arch_ok()` is always true.