[Alex0420W]

Updates v1.6 deferral Roadmap entry with all three identified blockers from multiple v1.6 attempts:

1. Polygon coverage (data; 3 of 9 polygons fit cleanly at 10m with the 100km horizon-march buffer)
2. File-resolver gap (closed in v1.8.3 by --lola-source flag)
3. Memory ceiling at 10m + 100km buffer (per-tile output array ~70 GB vs the GB10's 119 GB unified memory; OOM took the entire tmux session ~46s into the first tile)

Adds 'tile-internal chunking architecture' to 'Where this goes next' as the v1.9+ scope that resolves blocker 3 and unblocks 10m + finer analysis at the candidate-region geographic extent. Sketch: split each polygon-tile into ~16 sub-tiles of ~5500x5500 pixels with ~10000-pixel buffer overlap, process sequentially on GPU with mempool drains between sub-tiles, stitch outputs.

No methodology, data, or validation changes; v1.5/v1.7/v1.8 headline numbers (56/69 = 81.2%, median 1.69 km) unchanged.

Test count and CI behaviour unchanged (391 collected, 382 passed, 9 skipped — LEND + Robbins data-gated).

Summary

What this PR does, in 1–3 sentences.

Motivation

Why this change is worth making (link the issue if there is one).

Checklist

• ruff check . passes
• ruff format --check . passes
• pytest passes locally
• New behaviour has a test (or an explicit # noqa reason if not)
• Public-facing changes (CLI flags, criterion math, file layout) are
  reflected in the README or the relevant module docstring
• No real-data fixtures were committed; new tests run in CI without
  downloaded data (or are guarded with pytest.mark.skipif)

Scope notes

• Things this PR explicitly does not address (so reviewers know what
  to expect in follow-ups).