feat(master): score merge demand and record shadow placements

[beinan]

Part of #333 (P1.7 — last shadow-phase piece). Design §4.2, spec §2.3.

What

• scoring.rs (pure): merge_score = max(pending/min_gens, bytes/min_bytes) × (1 + wait/max_age). Classes: Critical (≥256 pending, or promoted for age ≥ max_turn_wait_secs), Normal (score ≥ 1), Tail. planner_order: class asc; class 1 oldest-first, never by score (review round 2); score desc elsewhere; name tiebreak → total order.
• Shadow pass in the planner, every reconcile: score → order → skip tails (real sweeps do too) → place on fittest executor by headroom → write Shadow assignments (replaced wholesale per pass, leader-token guarded). Then compare with reality (real MergeWal task active?) → scheduler_shadow_disagreements_total{kind} + per-pass gauge. This is the number that must approach zero before P2.
• Gauges: planner_merge_score{target}, planner_units{class}, planner_shadow_placements, planner_shadow_unplaceable.
• Spec test 5 as a source-level test: no file outside planner.rs/executors.rs/operator route references assignments.
• Deferred to P2: compaction/index scoring, class 2 commit-ready, max_consecutive_turns.

Validation

• Scoring: thresholds; age doubles score at max_age; promotion for age; class 1 oldest-first beats a 100× score (spec test 6); class before score; total stable order (20 items, forward == reverse).
• Planner etcd test extended: exactly one Shadow assignment for the 42-pending table on the live heartbeat, none for the empty table, headroom shows it, enqueuing a real task → next pass replaces not accumulates.
• Isolation test passes; clippy -D warnings, fmt, typos clean.

🤖 Generated with Claude Code