fix(inference,#1273): delete dead Candle Qwen3.5 GGUF backend (~1100 LOC)

[joelteply]

Summary

Delete the Candle-side Qwen3.5 inference path:

• inference/backends/qwen35_gguf.rs — Qwen35GgufBackend ModelBackend wrapper (194 LOC)
• inference/vendored/quantized_qwen35.rs — vendored hybrid DeltaNet + Attention recurrence loop (919 LOC)
• Module declarations in backends/mod.rs and vendored/mod.rs
• Replace the qwen3/qwen35 arm of backends::load_gguf_backend with a typed error pointing callers at LlamaCppAdapter

Net: +14 / -1132 LOC.

Why it was dead

AIProviderModule::register_adapters (modules/ai_provider.rs:221) only registers LlamaCppAdapter for local inference. CandleAdapter is imported but never instantiated. Qwen35GgufBackend was only reachable via backends::load_gguf_backend, whose only callers were unregistered (CandleAdapter, ContinuumModel, bin/* debug utilities) — none in the production hot path. Production Qwen3.5 chat goes through llama.cpp via LlamaCppAdapter → LlamaCppBackend.

Scope-down from initial #1273 plan

The original plan (#1273 (comment)) was to delete the entire Candle inference chain (CandleAdapter, ContinuumModel, quantized.rs, vendored qwen2/llama backends). cargo check confirmed broader scope is entangled with plasticity LoRA training tests, which use compact_llama_safetensors + model::rebuild_with_stacked_lora. That broader deletion needs a separate audit of plasticity's production reachability and is deferred to a follow-up card.

This PR keeps everything plasticity touches (model.rs, candle_adapter.rs, quantized.rs, llama_safetensors.rs, compact_llama_safetensors.rs, vendored qwen2/llama) and only deletes the Qwen3.5-specific Candle code that has no plasticity dependency.

Audit context

This is PR 2 of 4 for #1262 (silent CPU model fallbacks audit). Audit findings:
#1262 (comment)

Sibling PRs:

• fix(inference,#1262): delete dead compute_router.rs (no-CPU-fallback alpha) #1277 (PR 1) — delete compute_router.rs (213 LOC dead-code policy)
• perf(inference): delete vendored/metal_deltanet.rs (zero callers, misleading 'fall back to CPU' doc) #1274 (PR 3) — delete vendored/metal_deltanet.rs (zero callers)
• test(inference): regression test for no-CPU-fallback contract on LlamaCppBackend + ORT #1275 (PR 4) — regression test for no-CPU-fallback contract

Lane: alpha flywheel #1272 lane 6.

Wire change

backends::load_gguf_backend for qwen3/qwen35 architecture now returns a typed error ("Qwen3.5 GGUF routing through the Candle backend was removed in #1273. Use LlamaCppAdapter"). Callers in dead code paths (ContinuumModel, CandleAdapter) get a loud-fail rather than a silent dispatch to deleted code.

Test plan

• cargo check --features metal — clean (0 errors, 61 pre-existing warnings)
• cargo test --lib --features metal -- --test-threads=1 — 2096 passed, 0 failed (4 more than baseline; vendored qwen35 module removal cleared dead-code warnings that suppressed test discovery)
• Precommit (TypeScript + browser ping): PASSED

🤖 Generated with Claude Code