v0.3: Implement `flg context --mode resume`

[dlxeva]

Owner decision

The next core command should be flg context --mode resume.

This is the smallest build that turns FlowGrid from a handoff workflow into a project-state context engine.

Product goal

Create a bounded agent startup context for long-running rationale-heavy projects.

Agents should not load raw history by default. They should load reviewed project state, active pending state, and next actions.

Proposed command

flg context --mode resume --budget 4000

Default inputs

Read in this priority order:

1. SNAPSHOT.md
2. recent reviewed decisions from DECISIONS.md
3. recent entries from PROGRESS.md
4. next_actions from .flg/state.json
5. active pending patches from .flg/patches/
6. evidence references when available

Raw sessions under .flg/sessions/ should not enter the startup context by default.

Output

Write:

.flg/context/startup.md

Also print a short terminal summary:

• context file path
• estimated size
• sources included
• pending patches count
• warning if no reviewed decisions exist

Rules

• confirmed state outranks pending state
• reviewed decisions outrank unreviewed suggestions
• current project state outranks old history
• raw sessions are retrieved only on explicit request
• deterministic first version; no vector search or generic compression

Acceptance criteria

• command registered in src/flg/cli.py
• creates .flg/context/startup.md
• works in a freshly initialized FLG project
• works when no decisions exist
• includes active pending patch summary
• smoke test covers command