refactor(memory): use XML markers for injected context

[GCS-ZHN]

Summary

Replace the plain-text [MEMORY] header in injected context with explicit XML tags so the model can clearly distinguish reference material from the user's message.

Changes

• src/services/context.ts — wrap the injected block in <memory_context>...</memory_context> with a header reminding the model the contents are background context. Nest <user_profile> and <project_knowledge> sections; each memory item becomes <memory relevance="N%">...</memory>.
• src/services/user-profile/profile-context.ts — emit profile categories as <user_preferences>, <user_patterns>, <user_workflows>, and <learning_paths> with their items as nested elements carrying category / topic / frequency attributes. Add escapeXmlText / escapeXmlAttr helpers and apply them to all user-supplied strings so profile content cannot break out of the tags.

Behavior preserved

• formatContextForPrompt still returns "" when there is nothing to inject, so the if (memoryContext) guard in src/index.ts:279 behaves identically.
• getUserProfileContext still returns string | null; null is returned when no profile or no injectable items exist.
• The parts array name, forEach + parts.push calling style, and the workflow steps formatting expression are kept as-is. Only the pushed string contents change.

Motivation

Models that consume injected context benefit from explicit structural markers: tags make it trivial to identify the start/end of the injected region, attributes carry similarity / category / frequency without parsing prose, and escaping prevents profile content (some of which is written by AI cleanup passes) from breaking out of the wrapper.

Test plan

• No automated tests target formatContextForPrompt output today. Suggested follow-up: a unit test asserting the XML structure (see PR comment).
• Manual: trigger chat.message with a user that has a profile and project memories; confirm the injected part begins with <memory_context> and closes with </memory_context>.