feat(memory): strengthen default toolset instructions

[dgageot]

The default Instructions() in the memory toolset were written at a fairly high level of abstraction, making it easy for agents to leak tool usage into their replies or to skip recall when it would help. This PR ports the memory-handling lessons that were refined in the assistant/gordon.yaml doctrine and validated by the assistant/evals/memory-*.json eval suite into the shared toolset defaults, so every agent that uses the memory toolset benefits.

The changes tighten four areas. First, the silence rule now lists the specific phrases to avoid ("I'll remember that", "stored", "saved", "noted", "I searched my memory") rather than a vague directive, which makes it actionable for an LLM rather than interpretable. Second, a RECALL-first rule instructs agents to call search_memories before answering context-dependent or personal questions (e.g. "what's my name?"), with an explicit carve-out for greetings and self-contained informational questions so the tool is not overused. Third, a proactive STORE scan asks agents to check each user message for durable facts — including ones mentioned in passing — and to treat corrections as preferences rather than one-off overrides. Fourth, a secret-hygiene rule prevents storing secrets, tokens, or transient debugging details.

Two categories ("environment" and "correction") were added to the suggested taxonomy to match the new behavioral rules. Gordon-specific persona rules (banned words, response styles, empty-prose mechanics) were deliberately excluded; those remain in gordon.yaml where they belong.

go test ./pkg/tools/builtin/memory/... passes with the updated TestMemoryTool_Instructions assertions, and gofmt/go vet are clean.