refactor: extract MoE hybrid mode into common layer for qwen and laguna

[howard0su]

Move all hybrid infrastructure (placement, routing stats, storage, FFN eval, swap manager) from qwen35moe-specific files into server/src/common/moe_hybrid_* and add laguna hybrid mode with layer-by-layer decode.

Common abstractions:

• moe_hybrid_types.h: MoeHybridConfig, MoeLayerDesc (plain data, no vtable)
• moe_hybrid_placement: greedy knapsack expert placement
• moe_hybrid_routing_stats: runtime frequency tracking
• moe_hybrid_storage: hot/cold expert buffer management
• moe_hybrid_ffn_eval: GPU-resident, batched, single, prefill evaluation
• moe_hybrid_swap_manager: promote/demote at request boundaries

Laguna hybrid mode:

• init_hybrid_mode(): VRAM budget, placement, partial GGUF load
• hybrid_forward_one_token(): per-layer attn+router graph + hybrid FFN
• generate_hybrid(): prefill via laguna_step, decode via hybrid path
• load_target_gguf_laguna_partial(): skips expert tensor GPU upload

[cubic-dev-ai]

1 issue found across 102 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="harness/clients/run_opencode.sh">

<violation number="1" location="harness/clients/run_opencode.sh:14">
P2: `require_client_binary` uses a filesystem `-x` check only and does not fall back to PATH resolution (`command -v`), so valid PATH-based `OPENCODE_BIN` values that previously worked now fail.</violation>
</file>

<sub>Note: This PR contains a large number of files. cubic only reviews up to 100 files per PR, so some files may not have been reviewed. cubic prioritizes the most important files to review.
On a pro plan you can use ultrareview for larger PRs.

Re-trigger cubic</sub>

[cubic-dev-ai]

P2: require_client_binary uses a filesystem -x check only and does not fall back to PATH resolution (command -v), so valid PATH-based OPENCODE_BIN values that previously worked now fail.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At harness/clients/run_opencode.sh, line 14:

<comment>`require_client_binary` uses a filesystem `-x` check only and does not fall back to PATH resolution (`command -v`), so valid PATH-based `OPENCODE_BIN` values that previously worked now fail.</comment>

<file context>
@@ -11,6 +11,7 @@ source "$SCRIPT_DIR/common.sh"
 CLIENT_OUT="$LOG_DIR/opencode.out"
 EXPORT_OUT="$LOG_DIR/opencode-export.json"
 OPENCODE_BIN="${OPENCODE_BIN:-$CLIENT_WORK_DIR/clients/opencode/npm/bin/opencode}"
+require_client_binary "OpenCode" "$OPENCODE_BIN" "opencode" "OPENCODE_BIN"
 HOME_DIR="$LOG_DIR/opencode-home"
 PROJECT_DIR="$LOG_DIR/opencode-project"
</file context>