fix: reset checkpoint state on resume for fresh retry budgets

[jafreck]

Summary

When resuming a migration that previously failed (e.g. due to terminal exhaustion in Phase 4), the checkpoint retains stale failure markers that cause the resumed run to immediately re-fail or skip tasks instead of giving them fresh retry budgets.

This PR adds a prepareForResume() method that cleans up transient failure state on checkpoint reload, and fixes several related issues.

Changes

Checkpoint resume preparation (src/core/checkpoint.ts)

• prepareForResume() — new private method called on every checkpoint reload (resume) that:
  • Clears terminalExhaustion so the parity-gate loop re-enters instead of immediately raising TerminalExhaustionError
  • Resets failedTasks retry counters so code-migrator and parity-failure-resolver get fresh attempts
  • Clears blockedTasks so previously-blocked tasks re-enter the scheduling pool
  • Filters __phase4FlowCheckpoint.completedExecutionIds to only retain substeps for fully-completed tasks — failed/in-flight tasks re-enter from scratch (fresh code-migrator run, not just parity retry)
  • Resets both __flowCheckpoint and __phase4FlowCheckpoint status from 'failed' → 'running' so the Cadre flow runner re-enters correctly
  • Clears Phase 4 per-task cursor state for non-completed tasks

resetFromPhase() fix

• Reset flow checkpoint status from 'failed' to 'running' and clear error field so --from-phase works correctly with the Cadre runner

Copilot token usage fallback (src/core/agent-launcher.ts)

• Accumulate outputTokens from assistant.message events as a fallback when the usage summary block is missing from Copilot JSONL output

Agent prompt improvements (agents/templates/_partials/lore-index-first-principle.md)

• Stronger guidance to prefer Lore MCP tools over view/bash for reading source and target code
• Explicit tool name table so agents use exact MCP tool names

Test fixture config

• maxParallelAgents: 12 → 8
• resume: true → false

Testing

• Added 5 new tests covering resume preparation scenarios (terminal exhaustion clearing, failed task reset, flow checkpoint status reset, Phase 4 execution ID filtering, no-op for non-failed checkpoints)
• Updated existing terminalExhaustion persistence test to verify raw JSON instead of re-loading through CheckpointManager (which now clears it on reload)
• Added test for resetFromPhase flow checkpoint status reset