fix(verdict-cache): bounded per-entry LRU so a return to a recent state HITS (JEF-390)

[thejefflarson]

Closes JEF-390.

Problem

The verdict cache was single-slot per entry (cached: Option<(String, Verdict)>). A workload whose evidence oscillates between states A and B re-judged on every flip: judging B overwrote A's cached verdict, and flipping back to A missed — hammering the slow CPU model on a state it had already decided.

Change (engine/src/engine/state/verdict_store.rs)

• Replaced the single slot with a bounded per-entry LRU of (fingerprint → decisive Verdict) — a new VerdictLru (an ordered Vec with move-to-front + truncate; no crate).
• Default 32 slots, env-configurable via PROTECTOR_VERDICT_CACHE_SLOTS. Unset / unparseable / 0 → default; a positive value is honoured but floored at 2 (MIN_VERDICT_CACHE_SLOTS) so the cache can always retain one prior state. The cap is resolved once at construction, not per insert.
• cached_for hits on any fingerprint in the entry's LRU and promotes it to most-recently-used; cache_decisive inserts most-recently-used and evicts the least-recently-used past the cap.

Invariants preserved (no regression)

• Only DECISIVE verdicts are cached. cache_decisive is the sole insert path and the engine calls it only in its decisive arm; an Uncertain is never added and still arms the JEF-234 EntryBackoff (verified by test).
• Exact-fingerprint hit = byte-identical evidence (the fingerprint is the whole-prompt hash), so a served verdict is exactly as valid as when judged — wider retention introduces no new staleness.
• Journal-restore (JEF-141), display carry-forward, journaled dedup, recency (JEF-201), and the Mutex lock discipline are untouched. This ticket only widens the cached gate from 1 slot to N.

Tests

Added engine/src/engine/state/verdict_store_tests.rs (sibling file to stay under the 1,000-line cap), all driven through a small harness that mirrors the engine judging loop:

1. A→B→A serves A's verdict from cache on the return (2 judgements, not 3).
2. Beyond the cap the LRU evicts the least-recently-used state and re-judges it — and a revisit before eviction promotes the survivor (proves LRU, not FIFO).
3. Uncertain is not cached and the entry backs off (JEF-234); a later decisive verdict clears the backoff and populates the LRU.
4. The env knob parses: defaults on unset/invalid/0/negative, honours+trims positive values, floors 1 to the minimum.

Checks

• cargo fmt — clean
• cargo clippy --all-targets (warnings = errors) — clean
• cargo nextest run — 778 passed (4 new), file-size guard green

🤖 Generated with Claude Code

https://claude.ai/code/session_01VtjoJttCvBY4dzCoE4f9vP