feat: add reuseKb option to preserve knowledge base across fresh starts

[jafreck]

Problem

Starting a fresh migration (not --resume) always re-runs the knowledge-builder agent (Phase 2), which takes 12+ minutes for large codebases like zstd. When iterating on guidance, config, or migration strategy, the source code hasn't changed — the KB index and knowledge base are still valid and shouldn't need rebuilding.

Solution

New config option options.reuseKb (default: false). When true and starting fresh:

• Phase 0 (KB index): Already skipped via fingerprint guard — no change needed
• Phase 1 (task graph): Completion marker preserved from prior run
• Phase 2 (knowledge-builder): Completion marker preserved — skips the expensive agent call
• Flow checkpoint: KB step completions carried forward so the flow runner skips them
• Scaffold state: Preserved to avoid re-scaffolding

Everything else resets clean: planning (Phase 3), migration tasks (Phase 4), parity state, retries, token usage, and adjudication history.

Usage

{
  "options": {
    "reuseKb": true
  }
}

Then run without --resume:

npx aamf migrate -c migration.config.json

Logs: Fresh start with reuseKb — preserved phases [0, 1, 2], resuming from Phase 3

Changes

• src/config/schema.ts — Added reuseKb option with JSDoc
• src/core/checkpoint.ts — Extended load() to accept reuseKb, added loadPriorState() helper, logic to carry forward KB phases 0-2 while resetting everything else including flow checkpoint filtering
• src/core/runtime.ts — Pass reuseKb through to checkpoint loader
• tests/core/checkpoint.test.ts — 3 new tests: reuseKb with prior state, reuseKb without prior checkpoint, fresh without reuseKb
• tests/fixtures/zstd-c-project/migration.config.json — Enabled reuseKb: true

Test Coverage

• fresh start with reuseKb should preserve KB phases and reset everything else
• fresh start with reuseKb but no prior checkpoint should start from scratch
• fresh start without reuseKb should ignore prior state entirely

All 1505 tests pass.