feat(context): CAM Phase 2-C/2-D — LLM compress + embedding semantic scoring

[bug-ops]

Summary

• CAM Phase 2-C (feat(context): CAM Phase 2-C — LLM-assisted Compressed rendering via fidelity_compress_provider #4551): LLM-assisted Compressed rendering via FidelityConfig.compress_provider. When set, render_compressed() calls the named provider to generate a high-quality summary instead of truncating. The result is cached in MessageMetadata.deferred_summary and reused on subsequent turns. A 30 s timeout and a 2× token cost guard fall back to truncation silently on failure.
• CAM Phase 2-D (feat(context): CAM Phase 2-D — embedding-based semantic scoring to replace keyword_overlap #4552): Embedding-based semantic scoring via FidelityConfig.semantic_scoring_provider. FidelityScorer embeds the current query once per turn and computes cosine similarity against per-message embeddings cached in MessageMetadata.embedding (#[serde(skip)], in-memory only). Keyword-overlap remains the fallback when no provider is configured. w_keyword accepted as a deprecated alias for w_semantic.
• Docs (docs(scheduler): add terminology note distinguishing ContentFidelity (memory) from ContextFidelity (CAM spec) #4556): Disambiguation notes added to ContentFidelity (zeph-memory optical forgetting) and ContextFidelity (zeph-common CAM) to prevent confusion between the near-identical enum names.

Closes #4551, #4552, #4556.

Changes

• zeph-config: FidelityConfig gains compress_provider: Option<String> and semantic_scoring_provider: Option<String>; w_semantic accepts w_keyword alias
• zeph-llm: MessageMetadata gains embedding: Option<Vec<f32>> with #[serde(skip)]
• zeph-context: score_and_apply() is now async, split into embed_provider + compress_provider params; pre-pass embed loop; cosine_similarity + semantic_overlap functions; 30 s timeouts on all external awaits
• zeph-agent-context: call sites updated (await, separate provider handles wired from config)
• zeph-memory: embedding: None added to 3 MessageMetadata struct literals; ContentFidelity doc note added
• zeph-common: ContextFidelity doc note added
• Benchmark: fidelity_scorer.rs updated for new two-provider signature

Test plan

• cargo nextest run -p zeph-config -p zeph-context -p zeph-llm -p zeph-agent-context -p zeph-memory -p zeph-common — 3176 tests pass
• cargo clippy -p ... -- -D warnings — clean
• RUSTFLAGS="-D warnings" cargo check --all-targets — clean
• RUSTDOCFLAGS="--deny rustdoc::broken_intra_doc_links" cargo doc --no-deps — clean
• Live session test with compress_provider and semantic_scoring_provider set (blocked pending available provider in test env)

Follow-up issues (P3, not in this PR)

• No hard cap on embed/LLM compress input size
• LLM compress output not bounded by compressed_max_tokens post-processing
• Sequential embed loop (up to max_scored_messages sequential awaits; consider buffer_unordered)
• Benchmark covers fast path only; no bench for embed/compress paths