V5 Context-Subset Geometry Consistency Design

Decision

Add one opt-in training mechanism directly to the unchanged V5 evidence-residual model. The mechanism is named V5-CSGC (V5 Context-Subset Geometry Consistency). It is not a V6 revival and does not use the failed V6a localization supplier.

The preserved project contract is:

context RGB + calibrated cameras
-> V5 evidence-residual typed 3D state
-> doubled-resolution shared appearance
-> fixed zero-parameter renderer
-> target-camera measurements

The new path exists only during training. With context_subset_geometry_weight: 0.0, it is not executed. Inference, the V5 state dictionary, state construction, renderer, and target query path remain unchanged.

Motivation

V5 constructs geometry as:

base_density_logits = 4 * confidence - 2
completion_gate = floor + (1 - floor) * (1 - confidence)
density_logits = base_density_logits + completion_gate * density_residual

The existing rewrite penalty weights residual magnitude by confidence. It therefore constrains rewriting most strongly where evidence is already strong, while low-confidence regions retain the largest completion gate. Multi-scene telemetry shows that these completion residuals can approach their configured bound and vary substantially by scene. V5-CSGC tests whether stabilizing only the completion residual improves held-out geometry without replacing the V5 evidence mechanism.

The one-window V5 capacity result and the eight-scene result are different protocols. They are not treated as a same-protocol degradation curve. V5 remains the comparator until the frozen replacement gate passes.

Training Paths

For each four-context training batch:

1. The normal four-view V5 forward produces S4, all target-camera measurements, and the original supervised loss.
2. An independent deterministic generator selects one of the four context indices uniformly. The selected RGB image and camera are removed only after the dataset has established the original context-local coordinate frame and metric bounds.
3. The remaining three RGB images and cameras pass through the same V5 state encoder with shared weights and unchanged bounds, producing S3.
4. The three-view completion residual is inserted into the detached four-view evidence base:

z4 = b4 + g4 * r4
z3_to_4 = stopgrad(b4) + stopgrad(g4) * r3

This counterfactual isolates completion-residual instability. It does not punish deterministic evidence changes that are expected when one observation is removed.

Audited Overlap

Consistency is applied only where the subset path has at least two contributing context views and where deleting one view does not sharply change deterministic confidence:

shared = count(subset provenance > 1e-6) >= 2
stable = clamp(1 - abs(c4 - c3), 0, 1)
mask = stopgrad(shared * stable)
z_cf = mask * z3_to_4 + (1 - mask) * stopgrad(z4)

A zero-overlap batch therefore reproduces the detached teacher density and contributes finite zero consistency loss. It is never replaced with an unmasked loss.

Geometry Loss

The existing fixed renderer queries z_cf using the original target cameras. Only depth and visibility enter the new loss:

L_depth = visibility-weighted SmoothL1(depth_cf / voxel_spacing,
                                      stopgrad(depth_4) / voxel_spacing)
L_visibility = SmoothL1(visibility_cf, stopgrad(visibility_4))
L_csgc = L_depth + L_visibility

voxel_spacing is the mean metric cell edge after mapping stored [D,H,W] axes to metric [x,y,z] counts [W,H,D]. Depth is weighted by detached four-view visibility. Visibility is also matched explicitly so the subset path cannot reduce depth loss by making rays transparent.

RGB, high-resolution appearance residuals, point maps, and normals are not consistency targets. RGB would entangle appearance with the geometry hypothesis. Point is redundant with target rays and depth. Density-gradient normals are too noisy for the first mechanism test and remain an evaluation metric.

Gradient and Leakage Boundary

• The full-context geometry is stopped only inside the new consistency loss. It continues to receive the original V5 supervised and regularization gradients.
• The subset path receives the new consistency gradient through r3; its confidence, provenance, and all full-path evidence fields are detached in the new loss.
• Target RGB, depth, normal, point, visibility, and support never enter either state encoder.
• Target cameras remain renderer queries only.
• No subset-path target supervision is added.

Configuration and Compatibility

One strict training field is added:

context_subset_geometry_weight: float = 0.0

Nonzero use requires exactly four context views, state_architecture: evidence_residual, and mode: fixed. The effective weight ramps linearly from zero to its configured value over the existing optimizer-update warmup_steps; if warmup is zero, the configured value applies immediately. The first frozen proposal is 0.1 with the existing 3,000-update V5 warmup.

No learned module, model field, model parameter, checkpoint model key, or inference branch is added. When the configured weight is zero, the subset RNG, subset encoder call, counterfactual state, and consistency telemetry are skipped.

Required Telemetry

Each enabled JSONL logging step records the dropped and retained indices, full/subset evidence and gate means, shared-support coverage, stable overlap weight, residual disagreement, rendered depth drift in voxel units, visibility drift, residual absolute deciles, residual saturation fractions, and per-scene full/subset evidence, gate, and residual means. Lightweight scalar drift statistics remain available to the window/EMA tracker, while quantiles and per-scene audits are computed only when a JSONL record is due. The subset generator state is stored only in enabled-run checkpoints so an interrupted CSGC run can reproduce its deletion sequence.

Lower residual magnitude or lower training consistency loss alone is not success.

Acceptance Gates

Static and software gates

1. All pre-existing tests pass unchanged.
2. A V5 state dictionary loads with no missing or unexpected model keys.
3. Weight zero does not execute the subset path or alter V5 forward/training mathematics.
4. View deletion preserves the original camera matrices, bounds, order of retained views, and context-local coordinate frame.
5. Full teacher evidence and predictions receive no gradient from L_csgc; subset residuals do.
6. Zero overlap produces finite zero loss.
7. Full pytest, Ruff, source compilation, and an enabled synthetic CPU smoke pass before a real run.

Empirical gates

1. On the frozen one-window capacity protocol, PSNR may drop by no more than 0.20 dB and depth AbsRel and normal mean angle may worsen by no more than 5 percent relative to V5.
2. On held-out context deletion, geometry disagreement must improve by at least 20 percent and overlap coverage may fall by no more than 5 percent.
3. On the fixed eight-scene full-context validation, PSNR may drop by no more than 0.20 dB and depth AbsRel and normal mean angle may worsen by no more than 2 percent.
4. Replacing V5 additionally requires an absolute held-out geometry improvement. A lower training loss or lower path disagreement without better target geometry does not pass.

Failure Interpretation

• Better consistency with unchanged or worse held-out geometry indicates generic-geometry collapse.
• Better consistency with worse full-context metrics indicates that useful fourth-view evidence is being ignored.
• Falling visibility indicates opacity collapse even if depth disagreement improves.
• Falling overlap coverage indicates that the audit region is disappearing rather than stabilizing.
• Approximately two encoder forwards and one additional render per training step are expected; this compute increase is not interpreted as a model-capacity change.

No long run, validation tuning, or replacement claim is authorized by this design document.