fix(vis): custom __repr__ overlays lazy state on field section (closes mat#221)

[gerchowl]

Summary

Closes mat#221 — Bernhard's mat-vis#311 'Inconsistent outputs' sub-bullet. After v.textures triggers a fetch, repr(v) still showed every PBR field as None — no signal anything happened, even though the catalog data was loaded.

Pythonic decision: two namespaces, two purposes

The dataclass auto-repr was technically truthful (those fields are the override channel — user hadn't set them) but visually misleading. Two ways to fix:

Approach	Trade-off
Populate dataclass fields after fetch	Forces a magic distinction between "user-set" and "catalog-set" — needs an _overridden: set[str] to disambiguate, breaks replace / equality / pickle semantics, conflates inputs with outputs
Custom __repr__ overlaying lazy state	Field semantics unchanged; scalars= and available_textures= surface the resolved/cached view. Two clearly separate sections, no provenance ambiguity

We took the second. Same separation already pioneered by Vis.scalars (the public sparse view from the dispatch refactor): dataclass fields = caller overrides; property accessors = computed/resolved.

Output shape

Pre-fetch:

Vis(source='gpuopen', material_id='Aluminum Brushed', tier='1k',
    finishes={}, roughness=None, metallic=None, ..., fetched=False)

Post-fetch (catalog-backed):

Vis(source='gpuopen', material_id='Aluminum Brushed', tier='1k',
    finishes={}, roughness=None, metallic=None, ..., fetched=True,
    scalars={'metalness': 1.0, 'roughness': 0.4, 'color_hex': '#cccccc', 'ior': 1.5},
    available_textures=['color', 'normal', 'roughness'])

Post-fetch with caller override (v.metallic = 0.7):

Vis(source='gpuopen', material_id='Aluminum Brushed', tier='1k',
    finishes={}, roughness=None, metallic=0.7, ..., fetched=True,
    scalars={'metalness': 0.7, ...},   ← override won; visible in BOTH places
    available_textures=['color', 'normal', 'roughness'])

Implementation notes

• Touches self.scalars (catalog dict lookup) only when _fetched=True — unfetched repr stays cheap, no index IO triggered by print(v) or debugger inspection.
• Honors field(repr=False) on private cache slots (_textures, _fetched, _finish) — same fields excluded as the auto-repr.
• ~50 LOC total: 1 method on Vis, drops xfail decorator on 3 tests.

Test plan

• All 3 TestIssue221ReprObservability xfails flip to passing
• test_repr_does_not_leak_private_path (the existing repr contract test) still passes — class name + module unchanged
• No regression in 968-test full suite (vs 968 on main pre-fix; xfail count drops 11 → 8)
• Manual visual check across 4 cases (empty / overrides-only / pre-fetch / post-fetch with overrides) — all read clearly

Out of scope

Bernhard's remaining mat-vis#280-rooted error annotation gaps in test_error_messages.py (5 xfails) — separate concern, separate PR.

Refs mat#221, Bernhard's mat-vis#311