refactor(server): seal backend construction behind a plan

[Graffioh]

Summary

First PR in the server-refactor stack. It makes backend construction consume one model-aware, normalized plan and gives the resulting values stable, read-only groupings.

• make BackendArgs own operator input, including model and draft paths
• reduce server admission input to a narrow BackendAdmissionContext
• resolve GGUF metadata, placement compatibility, warnings, and SpecLA fallback once in prepare_backend()
• make create_backend(const BackendPlan&) the only backend-construction API
• make retained architecture and layer-split configs own their path storage
• group effective values into model, placement, cache, speculation, execution, and DeepSeek4 views
• use the same normalized values for construction, startup logging, tokenizer setup, cache policy, and /props
• keep ServerConfig and mutable BackendArgs out of backend construction

operator input + server admission facts
                 |
            prepare_backend
                 |
        read-only BackendPlan
      / model     placement \
 cache  speculation  execution  architecture options
                 |
     construction + server reporting

Ownership and boundary

Persistent backend configs use std::string for required paths and std::optional<std::string> for optional draft paths. Configs are moved into backend members, and .c_str() is used only at immediate C API boundaries. Constructed backends therefore do not borrow path storage from BackendArgs or BackendPlan.

BackendPlan is still the effective configuration source shared by construction and the server. This keeps normalization consistent without adding a second reporting snapshot. Such a snapshot should be introduced only if it can replace this dependency rather than duplicate the same state.

SpecLA environment mutation is deliberately unchanged: its graph helpers still consume process environment today, so removing that mechanism is a separate behavioral change.

Stack

This is PR 1 of 2:

1. refactor(server): seal backend construction behind a plan #691 — backend planning, owned configs, and grouped effective values (this PR, head 25fc625fd)
2. refactor(server): establish owned generation lifecycle #692 — owned generation lifecycle and runtime owner (head 2e6309477)

The history is intentionally two commits: c1a451fc0 establishes the sealed planning boundary and owned persistent configs; 25fc625fd groups the resulting effective values by concern.

#693 has been incorporated into this PR. The branches live in a fork, so both PRs target main; #692 is based on the previous head a567baef9 and needs a restack after this PR's review amendment.

Review only #692's follow-up delta: Graffioh/lucebox-hub@codex/backend-plan-boundary...codex/luce-engine-implementation

Verification

• compile-time ownership contracts cover every persistent backend and layer-split config
• backend-plan tests cover owned-path lifetime, construction API sealing, normalization, KVFlash fallback, unsupported-model fallback, and missing-draft failure
• feature-gate tests cover the server admission projection
• dflash_server, test_feature_gate, test_server_unit, test_deepseek4_unit, test_generate, and test_dflash build on HIP/gfx1151
• 449/449 model-free server and plan/feature tests pass
• git diff --check

Supersedes the docs-only proposal in #688 with executable boundaries and tests.

Review amendment

Consolidated SpecLA fallback cleanup in the plan builder, preserving warnings, DDTree selection, explicit thresholds, and environment actions. Added coverage for explicit DDTree settings on unsupported SpecLA models.

The feature-gate and backend-plan suites pass with both CUDA and HIP placement defines. The added case also passes against the pre-amendment implementation. git diff --check passes. These checks are model-free; no new GPU inference run was performed for this cleanup.

[cubic-dev-ai]

All reported issues were addressed across 12 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 12 files

<sub>Requires human review: Auto-approval blocked by 2 unresolved issues from a previous review of this commit.

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 12 files

<sub>Requires human review: Auto-approval blocked by 2 unresolved issues from a previous review of this commit.

Re-trigger cubic</sub>

[Graffioh]

@cubic-dev-ai review this PR

[cubic-dev-ai]

@cubic-dev-ai review this PR

@Graffioh I have started the AI code review. It will take a few minutes to complete.

[cubic-dev-ai]

Review completed against the latest diff

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>