P2-C: Full Metroid Recalc (DELIVERS: graph maintenance)

[github-actions]

Why: Incremental fast insert is approximate; need periodic full recalc. Recalc batch size must be bounded by H(t)-derived maintenance budget to avoid blocking the idle loop.

• P2-C1: Implement daydreamer/FullMetroidRecalc.ts

  • Query MetadataStore.needsMetroidRecalc(volumeId) for dirty volumes; prioritise dirtiest first
  • Load all pages in volume; compute pairwise similarities
  • Bound batch: process at most HotpathPolicy.computeCapacity(graphMass) pairwise comparisons per idle cycle (O(√(t log t)))
  • Select policy-derived max neighbors for each page; update MetadataStore.putMetroidNeighbors
  • Clear dirty flag via MetadataStore.clearMetroidRecalcFlag
  • Recompute σ(v) for affected nodes via SalienceEngine.batchComputeSalience; run promotion sweep

• P2-C2: Add Metroid recalc test coverage

  • tests/daydreamer/FullMetroidRecalc.test.ts
  • Test dirty flag cleared after recalc
  • Test neighbor quality improved vs fast insert
  • Test batch size respects O(√(t log t)) limit per cycle
  • Test salience recomputed and promotion sweep runs after recalc

Exit Criteria: Dirty volumes are recalculated in background within bounded compute budget; salience updated.