Light themes for dark-painted TUIs, and color-scheme reporting

[jamesyong-42]

Why

With a light ghosttea theme, several agent TUIs stay dark or half-broken:

• Grok paints #141414 on every cell, so the terminal background never shows. Its auto theme follows the macOS appearance, not the terminal.
• Claude Code's "Auto (match terminal)" theme reads our OSC 11 reply correctly, but never learns about a theme change, because we don't send mode-2031 reports.
• Codex and OpenCode already adapt from OSC 11.

A probe showed CSI ? 996 n going unanswered and DECRQM claiming mode 2031 support while no CSI ? 997 n was ever sent.

What changes

feat(terminal): report the color scheme

• The VT shim derives light/dark from the default background, answers CSI ? 996 n, and queues CSI ? 997 ; 1|2 n when a theme change flips the scheme and the program enabled mode 2031.
• New sessions get COLORFGBG (0;15 / 15;0) unless the caller set one.

feat(react): adapt dark-painted panes to light themes

• New renderers/light-adaptation.ts, applied per style in resolveStyle before blending. Glyph coverage, color glyphs and the shader stack are untouched.
• It only engages when the theme is light and at least 60% of a pane's cells carry an explicit dark background, and releases below 35%. Apps that follow the theme are never touched.
• Surfaces mirror their toe-corrected OKLab lightness around the theme background. Neutral ink keeps its distance from its cell background, and accents land in an APCA mid-tone band with their hue kept. Block elements use the surface mapping.
• TerminalTheme.lightAdaptation defaults to "auto" from configuration; embedders can pass "off" through the workspace theme override.

The rules come from capturing each app's dark and own-light screens through a private ghosttead and scoring candidate remaps against them. Mean ink ΔE against each app's own light design (lower is better):

Remap	Grok	Codex	OpenCode
Full-screen invert	5.0	10.1	5.5
This remap	3.1	3.3	3.7

A full-screen invert also flips hue: Grok's amber accent turns blue and Claude's red diff rows turn cyan.

Verification

• New tests: ghosttea-vt (996 reply, 2031 reports only on a scheme flip), ghosttea-core (theme change emits the report as WriteToTransport), ghosttea (COLORFGBG hint yields to the caller), light-adaptation.test.ts (parity with the measured colors, engagement thresholds, hysteresis, opt-out).
• ghosttea-react: tsc, eslint, prettier, and 275 vitest tests pass. The ghosttea-vt, ghosttea-core and ghosttea suites pass.
• Live daemon: with Claude Code's Auto theme highlighted, reloading the theme mid-session switches its preview to light. The previous binary stays dark.
• Real renderer: Grok frames replayed through WebGpuTerminalRenderer render light with the card border, gray hierarchy and amber accent intact.

service::protocol_tests::closing_a_session_records_why_it_ended is flaky here. It failed 1 of 4 runs on unmodified main too.

CI notes

• chore(deps) commit: Rust 1.99.0 (2026-09-28) added clippy::double_must_use, which fires on code generated by async-trait 0.1.89 in mesh.rs. That broke the required verify and windows gates for every branch. async-trait 0.1.92 fixes it; it and syn 3 need Rust 1.71, inside our 1.88 minimum.
• audit-published: expected to fail. This PR changes native sources, so the pinned Apple native artifact predates them until the next release rebuilds and pins it.
• dependency-audit: new npm advisories (vitest, electron, undici, brace-expansion). This is unrelated and fails on main too.

Not in this PR

• A Ghostty config key and settings toggle for lightAdaptation.
• Palette provenance in frame style records, which would let palette colors map exactly to the light theme's palette.
• A hue-preserving minimum-contrast.
• The remap in the Apple Metal renderer. iOS does get the signaling fix through the shared shim.
• A manual pass in the desktop app with Grok.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}