feat: inject active todos into model context after compaction and session resume

[platypusrex]

Problem

Todos persist in SQLite but are never re-injected into the model's context after two critical events:

1. Context compaction — todo_write/todo_read tool call messages get compacted away. The model loses all knowledge that todos exist. Nothing prompts it to call todo_read.

2. Session resume — the system prompt includes todo documentation (how to use the tools) but never the actual todo state. The model must independently decide to call todo_read, but nothing tells it active todos exist from a previous turn.

The todos are visible in the sidebar UI (use-session.ts loads them via getTodos), but the model is blind to them after either event.

Root Cause

• compactSimple and compactWithSummary in src/agent/compaction.ts operate purely on the Message[] array with zero awareness of the todos table
• buildInitialMessages in src/agent/index.ts constructs [system prompt, ...history, user message] without reading todo state
• getTodos(sessionId) is called in 4 places — all UI-facing (sidebar refresh). None inject into the model's context window.

Proposed Solution

A. Post-compaction injection (critical)

In src/agent/index.ts (~line 467), after messages.push(...result.messages), check for active todos:

const activeTodos = getTodos(args.sessionId).filter(
  t => t.status === 'pending' || t.status === 'in_progress'
);
if (activeTodos.length > 0) {
  messages.push({
    role: 'system',
    content: `You have ${activeTodos.length} active todos for this session. Call todo_read to review your task list before continuing.`,
  });
}

B. Session resume injection (significant)

In buildInitialMessages or the agent loop entry point, when building messages for a session with active todos, append a system message after history so the model knows to check.

C. Compaction preservation (optional, lower priority)

In shouldPreserve (compaction.ts lines 75-116), recognize todo_write tool calls as high-value messages worth preserving longer in context. This is a softer complement to A — the injection approach is more reliable on its own.

Scope

• ~30 lines across src/agent/index.ts and src/agent/compaction.ts
• A and B are the high-value changes
• C is nice-to-have

Context

Discovered during side panel todo overhaul work on fix/tui-issues. Related to TODO_ENFORCEMENT_SUGGESTIONS.md analysis — the persistence gap is the higher-impact problem compared to prompt enforcement for capable models like GLM-4.