Publish path densifies GeoGrid values() every cycle — sparse iteration follow-up to #21

[rolker]

Context

Part of #21 (live GeoMapSheet draft-tile persistence + Cartesian→geographic migration).

The migrated publish path (geoMapSheetToGridMap in grid_projection.cpp)
iterates each grid's cells via GeoGrid::values(), which returns a dense
std::vector<DepthAndUncertainty> in CellAreaIterator order (up to
960×960 ≈ 920k entries per grid), NaN-filled for empty cells.

Problem

On each publish cycle the node:

1. allocates a dense values() vector per grid (most cells typically NaN for a
   sparse survey patch), and
2. walks the full dense set, running residual per-cell lat/lon→ECEF trig only on
   the finite cells but still paying the allocation + full-length scan.

At GGGS scale on the live publish cadence this is heavier than necessary. The
per-cell ECEF conversion is also the residual cost the plan-review flagged once
the per-cell TF lookup was removed.

Proposed fix

• Provide / use a sparse accessor that yields only populated (CellIndex, DepthAndUncertainty) pairs, avoiding the dense allocation and full scan.
• Optionally hoist the ECEF conversion or precompute per-row factors where the
  longitudinal span is constant within a grid.

Acceptance

• Publish-path allocation no longer scales with total cell count, only with
  populated cells.
• A micro-benchmark or timed test showing reduced per-publish cost on a
  representative survey patch; published grid output unchanged (equivalence test
  still green).

---

Authored-By: Claude Code Agent
Model: Claude Opus 4.8