refactor(core): extract zeph-agent-feedback from zeph-core (Phase 1, PR 0b)

[bug-ops]

Summary

Extracts FeedbackDetector and JudgeDetector (1,933 LoC) from crates/zeph-core/src/agent/feedback_detector.rs into a new focused crate crates/zeph-agent-feedback/.

This is the first deliverable of the Phase 1 decomposition plan for #3480. It is the only file in zeph-core with zero Agent<C> coupling today, making clean extraction feasible without circular dependencies.

What changed

• New crate crates/zeph-agent-feedback/ with FeedbackDetector, JudgeDetector, and all associated types
• crates/zeph-core/src/agent/feedback_detector.rs removed; all 5 call sites updated to use zeph_agent_feedback:: directly
• No compatibility aliases or deprecated shims — all imports updated in place

Why this matters

• Creates the first real compile boundary: changes to feedback detection no longer trigger a full zeph-core rebuild
• Reduces zeph-core by 1,933 LoC (2.4% of total)
• Proves the extraction mechanism for the Phase 2 sibling crates (deferred until Agent<C> god-object decomposition unblocks)

Scope of Phase 1 vs issue acceptance criteria

Per the architecture investigation:

• AC#1 (30K LoC): Phase 1 (file splits + this extraction) will bring zeph-core to ~40K production LoC — the 30K target requires Phase 2 (new tracking issue to be filed)
• AC#2 (no file > 1,000 lines): addressed by subsequent Phase 1 PRs (test extraction, file splits)
• AC#3 (tests pass): all 8,623 workspace tests pass
• AC#4 (compile boundary): partially met by this PR (one subsystem isolated)

Test plan

• cargo build --workspace — clean
• cargo clippy --workspace -- -D warnings — zero warnings
• cargo +nightly fmt --check — clean
• cargo nextest run --workspace --lib --bins — 8,623 tests pass
• 127 unit tests in zeph-agent-feedback cover all public APIs (FeedbackDetector, JudgeDetector, CorrectionKind, JudgeVerdict)

Closes #3480