feat(skills,memory): skill evaluator, proactive exploration, compression spectrum

[bug-ops]

Summary

• research(skills): SkillLearnBench — continual skill learning benchmark for agent skill generation quality (arXiv:2604.20087) #3319 — SkillEvaluator: external critic LLM scores generated skills on correctness/reusability/specificity before disk write. Configurable weights/threshold; fail-open by default. Wired into all skill-producing paths (generator, proactive explorer, promotion engine).
• research(skills): reward-free spontaneous self-evolution — intrinsic meta-evolution without human supervision (arXiv:2604.18131) #3320 — ProactiveExplorer: classifies query domain (keyword-based) and generates a world-knowledge-{domain} SKILL.md when none exists. Runs before context assembly; new skill visible from the next turn (intentional MVP bound).
• research(memory): Experience Compression Spectrum — unified framework for memory/skills/rules at adaptive compression levels (arXiv:2604.15877) #3305 — CompressionLevel + RetrievalPolicy + PromotionEngine: budget-aware retrieval granularity and background episodic→skill promotion.

All three features are disabled by default (enabled = false); existing configs parse unchanged (#[serde(default)] on every new field).

Architecture

• crates/zeph-skills/src/evaluator.rs — SkillEvaluator, SkillQualityScore, SkillVerdict
• crates/zeph-skills/src/proactive.rs — ProactiveExplorer, DomainLabel
• crates/zeph-memory/src/compression/mod.rs — CompressionLevel, RetrievalPolicy
• crates/zeph-memory/src/compression/promotion.rs — PromotionEngine, PromotionInput
• crates/zeph-config/src/features.rs — SkillEvaluationConfig, ProactiveExplorationConfig, CompressionSpectrumConfig

Test plan

• cargo nextest run --workspace --lib --bins — 8326 tests pass
• cargo clippy --workspace -- -D warnings — clean
• cargo +nightly fmt --check — clean
• Live testing playbook: .local/testing/playbooks/skills-compression.md
• Coverage status rows added: skill-eval, proactive-explore, compression-spectrum (status: Untested)

Notable design decisions

• Proactive skill is visible next turn only — synchronous matcher rebuild on hot path would add an unbounded embed .await; next-turn rebuild is triggered by matcher_dirty flag.
• Evaluator defaults to fail-open: a critic error should not block skill generation; operators can set fail_open_on_error = false for strict enforcement.
• PromotionEngine::scan takes PromotionInput (not MemoryMatch) because MemoryMatch lacks session_id required for the min_sessions heuristic.
• MemoryError::Promotion with thiserror — no anyhow in library crates.

Closes #3319, #3320, #3305