#96: Resume Grok planner callbacks

[lukemaj]

Implements #96: Grok planner callbacks resume the saved planner session read-only instead of a fresh session.

• planner callback runs grok --resume SID -p PROMPT --verbatim --cwd WS --tools '' --permission-mode plan --no-subagents --disable-web-search --output-format json in the user's own Grok home (never a runner kit, no --always-approve); the answer counts only from the same session
• unresumable session falls back once to a fresh read-only session, recorded explicitly (invocation metadata grok_resume=fallback plus grok_planner_fallback ledger event); mismatch never falls back and blocks
• policy 2.8.0, regenerated skill reference, version 0.30.0 with changelog entry
• proof: python3 scripts/test.py 611 tests OK; live probe with grok 1.0.41 headless resume verified session identity

[lukemaj]

Rebased on main after #97 (0.30.0): resolved the Grok callback conflict by keeping #96's resume logic and passing main's new meta argument to both _finish_planner_callback calls; version now 0.31.0. Proof: python3 scripts/test.py, 627 tests OK (1 skipped).

[lukemaj]

Rebased on main e837872 (0.31.2) as 2671b9b; version 0.32.0 in VERSION and the three plugin manifests, #96 CHANGELOG entry on top of main's. Conflict resolution: kept #96's Grok read-only resume and fallback, main's meta argument on every _finish_planner_callback call, and regenerated skills/model-routing/references/codex.md from runner/policy.py (policy 2.8.0), which keeps main's escalation and one-PR delivery wording plus #96's resume wording.

Proof: python3 scripts/test.py locally: Ran 687 tests, OK (skipped=1). CI proof passed on 2671b9b: https://github.com/toolboxmd/model-router/actions/runs/36115983998

🤖 Generated with Claude Code