refactor(backends): self-describing WrappedServer backends (#2287)

[jeremyfowers]

Implements the plan in #2287: each inference backend describes itself with a plain-data descriptor plus a server class, and every scattered if (recipe == "...") site is rewritten to read a registry built from those descriptors.

What changed

Adding a backend is now one LEMON_BACKENDS line + a <stem>_descriptor.cpp (data) + a <stem>_factory.cpp (create()). No router, CLI, docs, or support-matrix edits — those are all derived.

• Descriptor types — BackendDescriptor / BackendOption / SlotPolicy (backend_descriptor.h); RecipeBackendDef moved to a shared header.
• Two-tier registry, generated from LEMON_BACKENDS at CMake configure time — a CLI-safe data registry (descriptors only, links into both lemonade and lemond) and a server-only factory registry (binds each descriptor to its class's create()). This split is what lets the CLI read recipe options/flags from descriptors without linking server classes.
• All 9 backends carry a descriptor (display name, binary, device, slot policy, options, support matrix, labels) + a create().
• Descriptor-driven sites (appendix rows 1–13): router creation, NPU/slot eviction & cloud LRU exemption (SlotPolicy), device type, recipe options / CLI flags / defaults, config-section identity, support matrix (RECIPE_DEFS deleted), recipe→label inference, cloud availability.
• /system-info recipes entries enriched with display_name / selectable_backend / uses_ctx_size / options / support. The desktop app now reads recipe display names from /system-info instead of hardcoded TypeScript.
• Docs generation — docs/tools/gen_backend_docs.py boots lemond, reads /system-info + server_models.json, and rewrites marker-delimited regions of docs/dev/backends-reference.md. A new CI job (backend-docs-drift) fails on drift. Authoring guide: docs/dev/adding-a-backend.md.

Corner cases / cleanups

• Unified ryzenai's config section to ryzenai everywhere (was inconsistent between s_backend_names and recipe_to_config_section).
• FLM↔exclusive-NPU eviction now keys off "NPU holder that isn't FLM" (correctly includes whisper-on-npu), fixing a latent recipe-string gap; collected-then-evicted to avoid iterator invalidation.
• Generic ModelInfo::extras map (from unknown server_models.json keys) so new backends add per-model fields without editing shared structs.

Verification

Local (this machine): lemond + lemonade CLI + web-app build green; tsc clean. Passing suites: server_endpoints (69), server_pinning (6), app-regression (37), test_model_name_normalization, test_cuda_arch_mapping. CLI run --help shows all descriptor-derived flags; /system-info carries the enriched fields; docs --check is clean.

Pre-existing failures unrelated to this change (reproduced identically on main): test_flm_status (stale message expectations, 16), test_llamacpp_system_backend (HIP plugin required on AMD-GPU hosts), test_multi_checkpoint_completeness (model pull/network), server_eviction (references phi-3-mini-4k-instruct-q4, absent from the registry), server_cli2 test_020 (built-in model name "Lite Collection" with a space breaks the test's whitespace parser). Relying on CI for clean-environment + cross-platform validation.

Notes for reviewers

• recipeOptionsConfig.ts (the deeply TypeScript-typed per-recipe option forms) is intentionally left to maintainers per AGENTS.md — the schema is now exposed via /system-info for a future dynamic migration.
• Backend install still goes through each backend's BackendSpec (install params are class-side behavior); the descriptor supplies the binary name.

🤖 Generated with Claude Code

[jeremyfowers]

CI status

All cross-platform builds pass (MSVC, AppleClang, GCC, Arch, openSUSE, Fedora rpm), validating the descriptor aggregate-init, CMake LEMON_BACKENDS codegen, and the CLI-safe/server-only registry split compile everywhere. Functional jobs exercising this change pass: CLI/Endpoints (ubuntu + macOS), Test .exe (whisper, moonshine, stable-diffusion, text-to-speech), backend-docs-drift, plus locally endpoints (69), pinning (6), app-regression (37).

The single red — Test CLI/Endpoints (windows-latest) → test_026_anthropic_messages_tool_calling — is a pre-existing flaky timeout, not from this PR. It's a 500 s ReadTimeout on a tool-calling inference request that the Windows runner intermittently can't finish in time:

• main run 27765794877: same job fails on the same test with the identical read timeout=500 signature.
• main run 27795912134: same job passes.

This PR touches backend construction, not inference, anthropic_api.cpp, or the tool-calling loop, so it can't change that test's latency. Re-running the job.