feat(preview): fine ROI renders + density contour -- the caves look like caves

Implements Codex's own recommendation from its critique, after it settled
the owner's two observations with evidence:

  "hard walls"  -> THE RENDERER. rg (room graph) re-rendered at step 1 is
                   rounded and organic. The hard edges were 4-voxel
                   sampling drawn nearest-neighbour. Fifth time in this
                   project that an alarming result was the instrument.
  "like a maze" -> REAL AND CORRECT. lt (lattice) stays rectilinear at
                   step 1 because it IS a maze; 19 of 64 candidates use
                   that source.
  "noisy"       -> ALIASING. nr#22 speckle 7 solid / 1 air at step 4 ->
                   0 / 0 at step 1, broad shapes intact.

Built: an optional fine ROI pass at step 1, radius 64 (~1.8M cells, cap
2M), run only for the 42 coarse survivors, plus a deterministic density=0
CONTOUR rendered beside the filled view.

Both views are kept side by side on purpose. The filled view is honest
about what was sampled and carries the walkability evidence; the contour
is legible. Codex's line, and it is the right principle: anti-aliasing the
coarse image "would only cosmetically blur the evidence and could invent
geometry". Do not replace evidence with beauty.

Fine labels carry their own sample step and resolved window so a fine
image can never be mistaken for a coarse one, and a MaxCells refusal shows
in the page rather than looking like a failed render.

  fine sampling 71.1 s · rendering 0.6 s · peak grid 9.34 MB
  42 filled pairs, 42 contour pairs, 0 refusals, 424 image links, 0 missing

Codex corrected my spec twice in the critique, both sharp: a binary step-1
raster ALWAYS has one-voxel axis-aligned edges, so only persistent
LARGE-SCALE right angles prove boxy generation (my criterion would have
produced a false "it is boxy" verdict); and persistent speckle is not
necessarily a roughness modifier.

Roughness deliberately untouched -- its call to judge that only after fine
rendering exists, and it stands.

⚠️ KNOWN, next: the fine ROI is centred on (0,0), so a candidate whose
geometry sits elsewhere renders blank (candidate 003 is a flat sheet with
no surface in the ROI). FVoxelStrateMetrics already carries
LargestComponentPoint -- centre the ROI there instead.

Build clean, 25 succeeded / 0 failed, verified by me.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VJiTcTSDNkKBtnNdiBtdaU

COMPOSER-NOTES.md

@@ -484,32 +484,46 @@ review a season before publishing.
 
 ### ✅ F1 measurement + preview built (2026-09-04)
 
-`FVoxelStrateSampleGrid` is an opt-in export of the exact `Air` polarity grid already sampled by
-`VF_MeasureStrate` / `VF_MeasureStrateWithSampler`. The editor/automation renderer consumes that
-buffer and never calls a density sampler, so each candidate's images and numbers come from one
-window and one sample pass. `VoxelForge.Composer.StructureRoll` now writes two PNGs per candidate
-and one self-contained `index.html` to:
+`FVoxelStrateSampleGrid` is an opt-in export of both the exact `Air` polarity and the scalar `Density`
+values already sampled by `VF_MeasureStrate` / `VF_MeasureStrateWithSampler`. The editor/automation
+renderer consumes that buffer and never calls a density sampler, so each candidate's metrics and images
+come from one window and one sample pass. The filled-cell view is paired with a marching-squares-style
+`density = 0` contour from the scalar field. The contour is not an anti-aliased or blurred coarse image:
+it is a legibility aid alongside the honest sampled-cell evidence.
 
-`Saved/ComposerPreview/structure_seed_0_corpus_9dbcea85_step_4_radius_256/`
+The step-4 measurement and 64-card contact sheet are unchanged in scope. Each card now shows the coarse
+filled and contour XZ/XY pairs. `VoxelForge.Composer.StructureRoll` writes them, plus an optional fine
+pass, to:
 
-The sheet contains all **64 candidates / 128 PNGs**, survivors first and then descending normalized
-distance from the corpus centroid so outliers are near the top. Every card repeats its exact resolved
-interior margin, sample step, XY bounds, and inclusive/exclusive Z span before listing air, largest
-component share, walkable fraction, feature scale, clearance, and arrival→departure. Rejected cards
-are visibly marked and name `vacuous`, `fragmented`, and/or `law failed`.
+`Saved/ComposerPreview/structure_seed_0_corpus_9dbcea85_step_4_radius_256_fine_step_1_fine_radius_64/`
+
+The fine pass is explicit and bounded: the current composer selects the 42 coarse survivors, samples
+step 1 in a radius-64 (128×128 XY) ROI, and refuses before allocation when `MaxCells=2,000,000` would be
+exceeded. The default is radius 64 because it stays around the observed 1.8M-cell budget while exposing
+voxel-scale detail; radius 128 is already roughly 7.1–7.5M cells and radius 256 exceeds the 8M coarse
+cap. The public settings let another editor/automation caller choose step, radius, centre, margin, and
+cap, or the writer can be called for an explicitly selected list. Fine filenames, captions, and each
+card's text include the sample step and exact resolved XY/Z window, so they cannot be mistaken for the
+coarse render. A refused fine render is shown as a refusal with its reason rather than silently omitted.
 
 The vertical image is an XZ section through the sampled cell nearest the strate-window centre Y. The
 plan image scans every sampled Z layer, keeps layers containing both air and solid, and chooses the
 one with the largest number of XY solid/air boundary transitions; ties prefer more mixed cells, more
 walkable cells, then the layer nearest the vertical centre. If no mixed layer exists it chooses the
 most solid fallback layer, so a slab does not silently produce a blank mid-height sheet. The focused
-run's first candidate was **128×48** for vertical and **128×148** for plan; dimensions vary with the
-resolved Z window. Both images use dark solid, light air, orange walkable cells, and a one-chunk
-footer scale bar. The page states **one chunk = 32 voxels = 8 m**, the density sign (`> 0` is air),
-the exact measurement window, the 512-pixel-per-dimension cap (including a 20-pixel footer), and
-that larger source grids aggregate already-sampled cells rather than resampling. It also carries the
-required caveat: a 2D apparent join is not a 3D connectivity proof; the separate arrival→departure
-verdict is the connectivity check.
+run's first candidate was **128×48** for coarse vertical, **128×130** for fine vertical, and **128×148**
+for both plan images; dimensions vary with the resolved Z window. Both views use dark solid, light air,
+orange walkable cells, a turquoise scalar contour, and a one-chunk footer scale bar. The page states
+**one chunk = 32 voxels = 8 m**, the density sign (`> 0` is air), every exact window, the 512-pixel cap
+(including a 20-pixel footer), and that larger source grids aggregate already-sampled cells rather than
+resampling. It also carries the required caveat: a 2D apparent join is not a 3D connectivity proof;
+the separate arrival→departure verdict is the connectivity check.
+
+The focused run measured **0.456 s** for the coarse render, **70.271 s** for the 42 fine scalar samples,
+and **0.573 s** for their filled/contour rendering. Peak fine `Air` + `Density` capture allocation was
+**9,338,880 bytes** and the peak single RGBA raster was **75,776 bytes**. The full StructureRoll test
+completed in **184.659 s**, with 64 coarse filled pairs, 64 coarse contour pairs, 42 fine filled pairs,
+42 fine contour pairs, and 0 refusals.
 
 ### 3.5 Validation in three layers, and how "good" is ever judged