(fix) bound per-session kernel-state and nudge-dedup caches with LRU eviction

[ranxianglei]

Fixes #113.

Problem
Two per-session caches grow without bound for the lifetime of the engine process:

• AcpStateStore.states (src/state.ts) — one full acp-kernel CompressionState per session id, including every block summary; delete() had zero callers.
• lastNudgeTurn (src/index.ts) — one entry per session id that ever received a nudge.

A long-lived server handling many sessions leaks memory proportional to total historical session count × per-session state size. (compressCallIdsToHide is intentionally left alone: it is bounded by in-flight compress calls and self-cleans on tool/result arrival.)

Cause
Plain Maps with only insert paths; nothing evicts idle sessions.

Fix

• New src/lru.ts: LruMap<K,V> (extends Map, API-compatible drop-in) + DEFAULT_SESSION_CACHE_LIMIT = 512. Recency refreshes on both get and set; over-cap inserts evict the coldest entry.
• AcpStateStore.states is now an LRU map capped at 512 sessions (constructor takes an optional limit for tests).
• lastNudgeTurn is now an LRU map with the same cap. Worst case after eviction of a cold session: one extra non-emergency nudge when it returns (per-turn dedup restarts) — bounded and self-healing.
• Why eviction is lossless for block identity: stateFor already rehydrates from the durable log (the same path a restart uses), and rehydration is deterministic — bN ids are recorded in the durable event or synthesised in ledger order, and message refs are re-derived from the append-only full-log array (allLogMessages includes shadowed originals), so numbering reproduces exactly.
• Secondary fix found while making eviction routine: rehydration set blocks/nextBlockId but not nextRunId, so the first compression after a restart (or eviction) reused run id r1 colliding with a rehydrated block. Rehydration now continues run ids after the rehydrated max (nextRunIdAfter).
• Kernel fields that do reset on eviction — tokenSnapshot, nudge cadence state, stats counters — all self-heal on the session's next turn; documented on the field.

Verification

• New tests:
  • tests/lru.test.ts (4): cap enforcement, recency refresh on get AND set, delete semantics, cap clamping. The re-set test caught a real bug in the first implementation (Map.set on an existing key does not move it to MRU position — fixed by delete+re-insert).
  • tests/state.test.ts "M2: store cap evicts cold sessions losslessly": cap-1 store, compress two distinct sessions, touch warm (evicts cold), then assert cold's rehydrated state has identical bN id, identical effectiveMessageIds, and continued nextBlockId/nextRunId — indistinguishable from the live state.
• npm run typecheck clean; npm test 213 pass / 0 fail (208 baseline + 5 new); npm run build clean (dist committed per AGENTS.md §4).

Built against main @ 6348af8 (v0.2.20).

[ranxianglei]

我来协助看一下这个问题,请分析并回复处理结果。