feat: MEMORY.md index file for memory discovery

[kokevidaurre]

Context

Claude Code uses MEMORY.md as a lightweight pointer file that's always loaded into the system prompt. It contains only links to memory files with brief descriptions — never memory content itself. This enables discovery without loading everything.

From Claude Code Source

• MEMORY.md: 200-line cap, 25KB cap, always in system prompt
• Contains only: - [Title](file.md) — one-line description
• No frontmatter, no content, no dates
• Truncated with warning if caps exceeded
• Separate from the topic memory files it points to

Current State

We have no discovery layer. run-context.ts reads state.md directly and injects it into the prompt. There's no way for an agent to know what memories exist without reading them all.

Proposed Changes

Add MEMORY.md at .agents/memory/{squad}/{agent}/MEMORY.md:

- [api-migration](api-migration.md) — Alembic pipeline state, rev 37 deployed
- [auth-rewrite](auth-rewrite.md) — OAuth2 rewrite for compliance
- [deploy-blockers](deploy-blockers.md) — Current deployment blockers

Reading behavior

1. If MEMORY.md exists, inject it into context (cheap: ~200 tokens for 10 pointers)
2. Agent decides which files to Read based on task relevance
3. If MEMORY.md doesn't exist, fall back to state.md (backwards compat)

Writing behavior

1. Agents write topic files, then update MEMORY.md index
2. SYSTEM.md instructions teach agents the two-step save pattern
3. Max 200 lines enforced by rotate script

Files to change

• src/lib/run-context.ts — load MEMORY.md instead of (or alongside) state.md
• .agents/SYSTEM.md — document MEMORY.md as index pattern
• scripts/rotate-states.sh — maintain MEMORY.md line cap

Expected impact

• ~200 tokens for index vs ~2000 tokens for full state.md
• Agent reads only relevant topic files on demand
• Enables per-turn relevance selection (test(workflow): fix child_process mock, add spawn export #710)

[kokevidaurre]

Closing — depends on #716 which was noise.