refactor(server): group backend plan by concern

[Graffioh]

Summary

Second slice of the server refactor stack, following #691.

• replace the flat BackendPlan::args() view with read-only snapshots by concern
• keep model path and inspected GGUF metadata together in Model
• separate Placement, Cache, Speculation, and generic Execution values
• put the four current DeepSeek-specific construction values in a small explicit DeepSeek4 group
• consume mutable BackendArgs once at preparation, after normalization and admission
• make backend dispatch and server startup/reporting read the same grouped plan
• remove the unused server-admission parameter from inert-feature warning collection

Execution now contains only stream_fd and chunk; it is no longer a miscellaneous model-path and DeepSeek bag. No architecture hierarchy or variant is introduced.

Stack

This is PR 2 of 3:

1. refactor(server): seal backend construction behind a plan #691 — backend planning and owned architecture configs, head d6ef2f7cf
2. refactor(server): group backend plan by concern #693 — grouped effective backend plan (this draft, single review commit 90062dc21)
3. refactor(server): establish owned generation lifecycle #692 — owned generation lifecycle, head 66b9295bb

This branch is directly based on the current #691 head. Because the branches live in a fork, the PR targets main and temporarily displays the preceding stack commits.

Review only this PR's delta: Graffioh/lucebox-hub@codex/backend-plan-boundary...codex/backend-plan-structure

The following #692 delta is: Graffioh/lucebox-hub@codex/backend-plan-structure...codex/luce-engine-implementation

Resulting boundary

BackendArgs (mutable input)
        |
  prepare_backend (consume once)
        |
BackendPlan {
  model,
  placement,
  cache,
  speculation,
  execution,
  deepseek4
}
   /          \
factory     read-only startup/reporting

The grouped field carriers are public read-only views; only the private builder can populate the plan. Persistent architecture configs introduced in #691 own their path storage, so the plan and returned backend have independent lifetimes.

Why this shape

• Model couples the path with the metadata inspected from that path.
• Execution remains small and architecture-neutral.
• DeepSeek4 is a simple field group, not a hierarchy or variant. A variant becomes useful only if more families accumulate distinct option sets.
• BackendPlan remains the one effective launch snapshot for construction and startup reporting.

A separate BackendRuntimeInfo is deferred until it can replace reporting state rather than duplicate it. Explicit SpecLA configuration is also deferred because removing the current environment bridge would be a broader behavior change across low-level graph helpers.

Tests

• HIP/gfx1151 builds: dflash_server, test_feature_gate, test_server_unit, test_deepseek4_unit, test_generate, and test_dflash
• 449/449 model-free server and plan/feature tests pass
• compile-time checks verify the direct ModelBackend return type and absence of a flat args() view
• focused tests cover owned model data and the explicit DeepSeek option group
• git diff --check

[Graffioh]

Incorporated into #691 as commit 6c9abb0. #692 is now rebased directly onto the combined #691 endpoint, leaving a two-PR server-refactor stack.