Fix custom provider icons in the model picker and Model Pad

[sambitcreate]

Summary

Fix custom provider artwork disappearing from the model picker and Model Pad.

• Recovers valid oversized stored PNG payloads during provider save by routing them back through native normalization instead of silently dropping them.
• Validates the re-encoded PNG against the renderer's bounded artwork contract before persistence.
• Preserves user-supplied PNG colors rather than converting them into monochrome masks; bundled provider glyphs remain theme-aware.
• Shows custom artwork on the Model Pad puck with a semantic, theme-resolved foreground for bundled monochrome marks.
• Bounds responsive Model Pad animation settling so infinite animations cannot hang hosted Electron tests.

Electron's getSize() and toPNG() already default to a 1.0 scale factor; explicit 1x arguments were removed because they were not the effective fix. The save-time recovery path is the behavior that restores previously oversized artwork.

Validation

• Exact-head CI is green at 391923f6, including the full verification and deterministic Electron E2E gates; the release-consumer contract is green.
• Real Electron E2E saves a valid oversized PNG through providers:save, verifies a decodable bounded persisted PNG, relaunches, and confirms the artwork survives.
• Focused provider-artwork tests, type-check, and lint pass locally at the current head.
• Three repeated artwork-persistence and responsive-layout Electron runs passed during readiness hardening.

Review status

• Rebased on current main, no merge commits, GitHub mergeable.
• Pullfrog is green and all review threads are resolved. Late causality feedback was addressed, and the scheduled-profile Electron helper was made deterministic at 391923f6 after it exposed the only failed rerun.

[sambitcreate]

Readiness hardening pushed at 8492109a after two independent adversarial reviews:

• preserve custom provider PNG artwork in original color instead of converting it to an alpha mask
• replace the Model Pad puck's hard-coded foreground with a resolved appearance token
• add a real Electron regression that saves a valid 128 px PNG through providers:save, verifies a decodable bounded 64 px result, relaunches, and confirms exact persistence
• bound Model Pad animation settling and ignore infinite animations in responsive E2E

Local verification:

• focused provider/UI/appearance tests: 46 passed
• Model Pad suite: 60 passed
• remote artwork projection + picker: 35 passed
• type-check, E2E type-check, lint, production build: passed
• provider-artwork and responsive Model Pad Electron E2E: 6/6 passed across 3 repeats each

Keeping the PR in draft until hosted exact-head CI is green.

[sambitcreate]

Hosted follow-up at fa526bee: the first exact-head E2E run reached the final 390×456 case as the test's global 90-second budget expired; the automatic retry completed the unchanged matrix in 24 seconds. I gave this exhaustive 24-state responsive case an explicit bounded 180-second budget without relaxing any geometry or reachability assertion.

Local verification after the change: E2E type-check and lint passed; the complete responsive Model Pad case passed 3/3 (1.2m).

[pullfrog]

ℹ️ No critical issues — one informational note inline, plus a description nitpick.

Reviewed changes

• Artwork persistence — parseProvider now persists through the new persistableProviderArtwork, which keeps contract-valid PNG artwork and re-encodes oversize payloads; the encoder re-validates its own output and fails closed.
• Picker / Model Pad rendering — ProviderIcon is refactored onto a shared ThemedProviderMark for bundled marks (custom artwork still renders as an original-color <img>), and the Model Pad puck now renders the active model's provider icon.
• Knob theming — new --model-pad-knob-foreground token resolved from foregroundForFill("#FFFFFF", scheme) with static fallbacks, applied to .model-pad-knob and its confirmed state.
• Coverage — new source-reading provider-artwork suite registered in test, test:coverage, and test:slash-commands, expanded composer/picker/display tests, a new Electron E2E that saves a 128px PNG, relaunches, and asserts a bounded 64px persisted result, and a bounded Model Pad responsive case.

ℹ️ Nitpicks

• The PR description still says custom marks render "with the same currentColor mask as bundled logos," but the final code renders them as an original-color <img> and .papercuts/troubleshooting.md documents that reversal. Worth updating so the merge record matches the implementation.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using DeepSeek Flash ｜ 𝕏}

[sambitcreate]

Addressed the late review at 0e31ee99: removed the no-op explicit getSize(1) / toPNG({ scaleFactor: 1 }) changes, rewrote the source contract and E2E naming to credit the actual save-time recovery, and corrected the PR/papercut narrative. Focused artwork tests, type-check, and lint pass locally; fresh exact-head CI is running.

[pullfrog]

✅ No new issues found.

Reviewed changes

• Reverted the no-op 1x encode path — removed PIXEL_SCALE and the explicit getSize(PIXEL_SCALE) / toPNG({ scaleFactor: PIXEL_SCALE }) arguments, restoring the base getSize() / toPNG() calls; createFromBuffer(source.bytes, { scaleFactor: 1 }) is unchanged from base.
• Retargeted provider-artwork.test.ts — dropped the assertions on the removed encode arguments and now pins the post-encode normalizeProviderArtwork check plus the persistStoredProviderArtwork recovery wiring.
• Corrected the narrative — .papercuts/troubleshooting.md now credits save-time recovery instead of the Electron-default 1x arguments, and the artwork E2E is renamed from Retina to Oversized.

Verification at this head: focused provider-artwork suites 7/7 pass, tsc --noEmit and type-check:e2e clean. The prior review's inline thread on the no-op encode is resolved, and the PR description now matches the implementation.

  ^{｜ View workflow run ｜ Using DeepSeek Flash ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes

Reviewed the single commit added since the prior Pullfrog review (391923f6), a test-only stabilization of the Assistant/Scheduled E2E.

• Stabilized scheduled-search cleanup — the E2E now clears the controlled scheduled-task search through the native HTMLInputElement.value setter plus a bubbling input event instead of an empty Playwright fill("").
• Documented the quirk — a papercut records that empty fill("") can leave a controlled search unchanged on hosted Electron and points at the native-setter workaround.

The replacement is the canonical React controlled-input workaround: Playwright's fill for input[type=search] selects the value and calls document.execCommand("insertText", false, value), where an empty string can be a no-op; calling the prototype setter bypasses React's per-node value tracker so the dispatched input event registers as a real change. The follow-up toHaveValue("") and No matching tasks assertions still prove React state cleared, so the workaround cannot mask a product regression. npm run type-check:e2e is clean at this head; the deterministic Electron E2E was still running when this review was written.

  ^{｜ View workflow run ｜ Using DeepSeek Flash ｜ 𝕏}