fix(catalog): align AMD395 Linux local-model parity

[qinyue0508-cmd]

Summary

• align the Linux AMD395 HIP startup health timeout with the amd395-win baseline (180s) without modifying the Windows branch
• lock platform-specific HIP resolution across all 13 catalogued llama.cpp GGUF models
• cover local LLM, VLM, and embedding parity while preserving Qwen3.6 BF16/safetensors native resolution
• keep the PR diff to 3 files after rebasing current develop and synchronizing current amd395-linux

Final head and artifact

• final head: 4ea8f7a42688020ff8aa538cbaf956a762513925
• artifact: aima-linux-amd64-v0.5-dev-amd-strix-halo-4ea8f7a42688
• SHA256: 1f58495be3202932363645dd5f99d1099d8456f39bac1e2c0a7aee301c3a6002
• build-metadata.json.git_commit: 4ea8f7a42688020ff8aa538cbaf956a762513925

Automated validation

All commands passed on the final head:

go test ./... -count=1
go vet ./...
go test -race ./internal/engine ./internal/knowledge ./internal/hal ./cmd/aima -count=1
make amd395-linux-build-test

The parity matrix resolves all 13 GGUF assets on both AMD395 platforms: 10 LLM, 2 VLM, and 1 embedding model. Windows selects llamacpp-hip-windows; Linux selects llamacpp-hip-linux.

Baiying AMD395 evidence

Validated from the same final artifact with an isolated AIMA_DATA_DIR and unused ports 38251-38255.

• HAL: Radeon 8060S, RDNA3.5, gfx1151, 98304 MiB unified memory, ROCm 7.2.3
• full native 1.5.1 bundle import: nested .../aima-engine-native-amd395/1.5.1/bin/aima-engine visible in scan/list/info
• imported native version: detected 1.5.1, catalog 1.5.1, exact, available=true
• with AIMA_ENGINE_DIR unset, imported aima-engine --version and deploy both resolve from isolated dist
• single-file launcher import copied configured companion directories; imported launcher independently returned aima-engine-native 1.5.1-native and scanned exact/available
• negative gate: 1.4.1 scanned as mismatch / available=false; deploy exited 1 before opening its port
• BF16/safetensors: aima-engine-native-amd395, health OK, ready/running, chat returned AIMA Linux BF16 OK
• UD-Q4 GGUF: imported llama.cpp b9637 scanned compatible / available, resolved llamacpp-hip-linux, HIP offload used --n-gpu-layers 999, health OK, chat returned AIMA Linux GGUF OK
• explicit-path precedence: DB=GGUF + explicit safetensors resolved native; DB=safetensors + explicit GGUF resolved llama.cpp/int4
• cleanup: deployments stopped, ports 38251-38255 closed, test processes absent, isolated test directory removed

@rjckkkkk please review.

[rjckkkkk]

Reviewed the catalog timeout and AMD395 Linux/Windows GGUF resolution changes. CI and local verification pass; approving for amd395-linux.

[rjckkkkk]

已批准并合并到 amd395-linux。

合作方下发请使用合并提交对应的分支构建包，不要继续使用 PR head 包：

• Merge commit: ddac975601aad8f9c19308a4b76cee67a654ce46
• Workflow: AMD395 Linux Build #15
• Artifact: aima-amd395-linux-ddac975601aa
• Binary: aima-linux-amd64-v0.5-dev-amd-strix-halo-ddac975601aa
• Binary SHA-256: 604f4934aa57fa86ec4e4732f55a0cbafe641149b3d4702140e715cbf940adfc
• Artifact expires: 2026-09-10 08:29 UTC

CLI 下载：

gh run download 31473297340 --repo Approaching-AI/AIMA --name aima-amd395-linux-ddac975601aa --dir ./amd395-linux-ddac975601aa
cd ./amd395-linux-ddac975601aa
sha256sum -c checksums.txt

已复核 build-metadata.json.git_commit、Linux/amd64/AMD Strix Halo 目标信息和二进制 SHA-256；这是开发候选包，不是正式 GitHub Release。合作方部署前请继续按 docs/amd395-linux-build.md 在 AMD395 Linux 真机完成干净数据目录和最小推理验收。