Flip the manifest repair-layer default for v1 element kinds from
`light` to `system`, so weak-model designs adapt to the document's
light/dark axis instead of baking light-mode hex. Both prerequisites
are now in place: the orchestrator seeds the 56-token semantic palette
at run start (`-*` refs resolve at paint via scene_vars) and
every v1 builder emits numeric gap/padding in system mode. Explicit
`theme=light|dark` still wins.
ab-v9.3 validation (52 prompts × 4 arms, theme=system default):
- Rendering: 0 dangling $color-* refs across all 208 .op; 207/208
carry the seeded palette + Mode:[Light,Dark] axis. The D2 grey-ref
risk does not materialize at corpus scale.
- Structural inertness: theme is injected post-model by repair_args
and only selects fills; role inference + the contrast post-pass are
already $ref-aware (role_post_pass resolves a ref to None and skips).
M3/M5 cannot move on the flip — glm's 15 M3 misses are 100%
missing-role (model output), theme-independent.
- No collapse: M3 ark 81% / glm 71% / DS 85% / m3-think 94%, 0-1
garbage per arm. Per-arm deltas vs ab-v9.2 are cross-codebase
(Kayshen's d31734e2 edited the manifest skill + builders between
runs) + sampling, NOT the theme flip.
Two failures surfaced once the casement build blocker cleared and the
`cargo test --workspace` step finally ran on v0.8.0-new:
- wheel_over_panel_scrolls_rows_and_clamps: `over_topmost_panel` ran
first in apply_wheel and lists the variables panel, so a wheel over
the panel was swallowed (returned true) WITHOUT ever reaching
try_scroll_variables_panel — rows never advanced. Move the variables
scroll handler ahead of the topmost-panel guard; it already swallows
the wheel over its own rect, so the canvas-zoom guard still holds.
- variables_panel_preset_button_toggles_menu: the second press used a
+58 y-offset that lands in the open dropdown's SaveCurrent row, which
intentionally focuses the name input and keeps the popover open (TS
parity, see anchor_press_does_not_double_toggle). Re-press the button
itself (+22) so the test exercises the toggle it is named for.
op-host-native lib: 296 passed / 0 failed.
theme=system audit (D2 prerequisite): the schema accepts $spacing-*
expression refs in gap/padding, but both Rust layout paths zero them
out — jian-core container_to_style and op-pen-loader gap_value fall
through Expression to 0 — so a ref collapses the layout instead of
resolving. Exactly three emission sites existed: modal_shell's card
padding ($spacing-5) and gap ($spacing-3), and toast's pill gap
($spacing-2). Emit the palette-equal numbers (24/12/8) in every mode,
same pattern as the numeric font-size toast fix (011f4104). A new
catalog-wide invariant test builds all kinds with theme=system and
rejects any expression gap/padding, so future ports can't reintroduce
the class. Color $refs are unaffected — scene_vars resolves fill and
stroke refs at paint.
Replace the global OPENPENCIL_MANIFEST env gate with per-model routing:
manifest defaults ON for the families that cleared the ab-v9.2 KPI gate
(minimax 92% / ark-code 92% / glm 98% / deepseek 83%, all >=70%) and
stays OFF for models without benchmark data — the catalog is a floor
for weak models, not a ceiling for strong ones. The env var remains a
both-ways override (1/true/on forces on, 0/false/off forces off) so
op-smoke benchmarks and rollback keep their one-knob workflow.
Also plumb the selected built-in agent's model id into DesignRequest:
the production chat path always sent model:None, which routed every
model to Full-tier prompts and dead-armed both this gate and the M3
thinking policy outside op-smoke. Live-verified: glm-5.1 with no env
set engages the manifest protocol purely via routing.
Port the TS applySemanticPalette 56-token palette (22 themed light/dark
colors on a Mode axis, 6 chart singles, 28 numerics) and wake the
dormant variables.rs seeding path. Merge semantics differ from TS:
editor-core set_variables_bulk extends (incoming wins) while TS keeps
existing values, so the snapshot computes the missing set first and
seeds only absent tokens via one MergeThemePreset; rollback removes
exactly the created names/axes when no content survives. Live-smoke
verified: generated .op carries the 56 vars + Mode axis and $color-*
refs resolve at paint.
A childless Frame with explicit stroke or fill is pixel-equivalent to a
childless Rectangle, which the blank check already admits. otp_input's
empty stroked digit slots made perfect manifests read as blank
scaffolding -> retry ladder -> raw fallback degradation (ark emitted
flawless otp manifests twice and got rejected both times). Also
serialize the OPENPENCIL_MANIFEST env tests behind a shared lock --
parallel set/remove races only surfaced in full-suite runs.
ab-v9.1 misses were attention failures, not vocabulary failures: kinds
named literally in the brief (toolbar 5/10) still got hand-composed,
while anti-pattern-named kinds rarely missed. Token-match the subtask
label+elements text (plus the whole brief for single-section plans)
against the catalog, expand via a data-grounded synonym table, and
inject up to 6 nominations as an ELEMENT HINTS block with composite-
first and no-hand-compose rules. ab-v9.2 matrix: 51 FAIL->PASS vs 4
noise-shaped regressions; all four arms >=83% M3-pass, M5 adoption 96%.
Full-tier prompts run ~8k tokens and deepseek kept hand-rolling past a
top-of-prompt element catalog (ab-v9 58% vs Basic-tier 85%). Pull the
element-manifest skill out of priority order and append it last, right
above the output contract, so recency keeps the catalog in attention.
The skill's COVERAGE section now also teaches composite-first checking
instead of only naming examples.
M3 with thinking disabled drops to M2.7-level element adoption; route
design sub-agent calls through Adaptive thinking with a 16384 budget
for the M3 family while other built-in models keep Disabled+8192.
Validated by the ab-v9.2 matrix (m3-think arm 92% M3-pass).
check-widget-boundary.sh F4 forbids op_editor_ui::widgets references
outside widget_host.rs + siblings; the remote-icon fetcher now calls the
new icon_ingest wrapper instead of icon_catalog directly.
