fix(memory): preserve native reranker score scale - #4249

Summary

• Stop clamping memory-plugin reranker scores and thresholds to [0, 1], so logit-scale rerankers can use native negative or >1 scores.
• Apply the behavior consistently across shared recall logic and Claude/Codex/OpenCode/DSH/PI/ZCode integrations.
• Update configuration docs and add regression tests for native-scale thresholds.

Root cause

The memory plugin path assumed reranker scores were cosine-like [0, 1] values. Local OpenAI-compatible rerankers such as bge-reranker-v2-m3 can return raw logits, where relevant results may be negative. Clamping both configured thresholds and item scores made valid operator settings like -8 ineffective and could filter out every recalled memory.

Fixes #4224.

Test plan

• node --test examples/memory-plugin-shared/recall-core.test.mjs examples/memory-plugin-shared/agent-hook-runtime.test.mjs examples/opencode-plugin/tests/config.test.mjs examples/pi-coding-agent-extension/tests/config.test.mjs examples/dsh-memory-plugin/live-recall.test.mjs examples/codex-memory-plugin/scripts/config.test.mjs examples/claude-code-memory-plugin/scripts/config.test.mjs examples/codex-memory-plugin/scripts/auto-recall.test.mjs examples/claude-code-memory-plugin/scripts/auto-recall.test.mjs
• git diff --check

Note: I also ran the broader memory-plugin PR test command after npm ci --prefix examples/dsh-memory-plugin. The #4224-related tests passed; unrelated local installer tests still fail in this checkout because the dev installer resolves the marketplace directory as /examples, and local TRAE CLI detection now includes trae-cn where existing assertions expect only trae.