feat(tui): the quiet-constellation skin — §6.1 art direction lands [spec 10]

[wine-fall]

Implements specs/spec10/10-tui.md §6.1 (art direction) — composition notes written back into §3.3 and §3.6.

What this is

The §6.1 creative session picked its skin: 04-quiet-constellation (from scratch/ui-concepts/logo-pet, an untracked concept set). This PR realizes it without reopening the §3.3 contracts — same four regions, new composition and light:

• Palette (tui/src/palette.ts): one deep blue-charcoal night ground shared by every scene; each scene is a single near-monochrome accent family (dawn cream / daylight sage / ember / moonlight). The hour changes the warmth of the light, never the room.
• Constellation renderer (tui/src/constellation.ts, new, pure): a seeded starfield, the existing 24fps FFT bins re-rendered as a floating braille particle mist (frayed top, sparse fallout under the baseline), and the pet blitted into a cleared halo mid-sky. Deterministic by construction — seeded PRNG + hashed jitter — so a resize never reshuffles the sky and the arithmetic is unit-assertable.
• Layout (tui/src/app.tsx): one breakpoint. At ≥96 columns the alive band recomposes as a sky panel beside a poem-spaced log, with now-playing as a dim ♪ line under the panel; narrower terminals keep the classic bottom band unchanged. The sky paints on its own 12fps clock — viz frames only feed the smoother — so stars twinkle and the pet animates through silence.
• Engine/wire: zero changes.

Verification

• 10 new unit tests on the renderer (dimensions, seed determinism, silence, level-monotonic dot count, glyph set, pet blit, hostile bins); full gate green: 661 passed, typecheck + oxlint clean.
• Visual check against the concept mockup: the composition was rendered frame-for-frame from the real modules and compared side-by-side with 04-quiet-constellation.png, then tuned (floating baseline, mist density, pet position) until the structure matched.
• Real-boundary smoke (STUB brain, real pty): 80×24 boots the narrow band; 130×38 boots the sky panel (star glyphs + pet half-blocks painted, no braille in silence — correct), no crashes, clean teardown.

Peer review (codex gpt-5.5 xhigh): 1 finding, 0 applied — the one finding was mechanical (the new files were untracked at review time, so its diff view flagged a "missing module"; the commit itself resolves it), 0 genuine defects.

By-eye tuning of the new skin in a real terminal stays issue #79 (the standing §6.1 acceptance pass), now with concrete knobs to judge: mist density/baseline, star density, twinkle rate, pet position, per-scene accents.

AI coding brief

Original request: A TUI redesign session: four concept mockups were compared for terminal feasibility; the user rejected the analog-radio layout as ugly, narrowed to concepts 02/04, and delegated the final pick — 04 won because its atmosphere is procedural (code, fast to iterate) rather than hand-drawn asset work, its quiet sparseness fits murmur's identity, and it needs zero engine changes.

Manual interventions: Mid-build the user added one hard requirement: compare the result against the concept screenshot instead of free-styling. That produced the render-to-HTML comparison loop, which caught the first version's real miss (a dense bottom-glued bar texture instead of a floating mist) and drove two tuning rounds.

Retro: The self-grilling round before building surfaced a contradiction in the original plan (a fixed amber monochrome would have killed the already-built scene tinting) and killed three scope items cheaply. Asking for the mockup-comparison loop up front, rather than mid-build, would have saved one round trip.

🤖 Generated with Claude Code