diff --git a/README.md b/README.md
index e5d359501..b1162fff9 100644
--- a/README.md
+++ b/README.md
@@ -20,6 +20,7 @@
+ 
@@ -34,13 +35,18 @@ Lucky is a **production-grade, self-hosted Discord music bot** — built as a Ty
Unlike Groovy, Rythm, or Hydra (all shut down by third-party enforcement), Lucky can't be taken offline because **you host it.** Every feature is included — no paywall, no premium tier.
-- **Live demo**: [lucky.lucassantana.tech](https://lucassantana.tech)
+- **Live demo**: [lucky.lucassantana.tech](https://lucky.lucassantana.tech)
- **Invite to your server**: [Add Lucky now](https://lucky.lucassantana.tech/invite?utm_source=github&utm_medium=readme&utm_campaign=readme-badge)
---
## ✨ Features
+
+ 
+
Everything most servers need, in one bot — see it live.
+
+
### 🎵 Music Player
- **Multi-source playback**: YouTube, Spotify, SoundCloud
- **Smart autoplay**: Recommendations based on your listening history
diff --git a/assets/screenshots/landing-features.png b/assets/screenshots/landing-features.png
new file mode 100644
index 000000000..144b65e45
Binary files /dev/null and b/assets/screenshots/landing-features.png differ
diff --git a/decisions/2026-07-12-brand-asset-regen-tooling.md b/decisions/2026-07-12-brand-asset-regen-tooling.md
new file mode 100644
index 000000000..d2d1214df
--- /dev/null
+++ b/decisions/2026-07-12-brand-asset-regen-tooling.md
@@ -0,0 +1,171 @@
+# Brand-asset regeneration — hybrid resvg-frame + Gemini-mascot via inference.sh
+
+- **Date:** 2026-07-12
+- **Status:** Accepted — pilot completed (see Pilot outcome + FINAL DECISION below: keep the original mascot)
+- **Deciders:** Lucas Santana
+- **Method:** `/research-and-decide` — CLI-verified capability probe + web-doc research + artifact-only `decision-critic` (returned NEEDS_REVISION → decision revised to hybrid)
+
+## Context
+
+The README hero `assets/lucky-social-preview.png` (1280×640 baked raster) is off-brand: a
+neon-glow cat mascot with **gold/orange** stripes + eyes, plus overlaid text chips
+("Lucky", "TypeScript", "React", "849+ Tests", "4 Packages") where some accents render gold.
+Brand palette is Discord **Blurple `#5865f2`** + **Neon Pink `#ec4899`** (gold explicitly
+removed — `packages/frontend/branding/BRANDING_GUIDE.md`). No source pipeline composes this
+PNG (repo grep: referenced only as a static file in `README.md` + `docs/TOP_GG_SUBMISSION.md`).
+
+Scope corrections established during research (do NOT route these through AI):
+
+- **Production `og-image.png`** (served to web) is a SEPARATE asset, already code-generated at
+ build time via **resvg** from SVG in `packages/frontend/scripts/prerender-seo.ts` — clean,
+ brand-colored, no gold. Out of scope.
+- **`assets/lucky-logo.svg`** is off-brand **purple** (`#7c3aed`/`#8b5cf6`/`#a78bfa`/`#c4b5fd`)
+ → fix by deterministic fill-swap, not AI. Separate cleanup.
+- **`assets/outline-layer-{all,mid,gold}.svg`** are orphan design exports, referenced nowhere
+ → ignore (YAGNI).
+- **`--color-lucky-gold-*`** CSS vars (`index.css:130-132`) are legacy compat aliases already
+ holding pink values, 0 live refs, slated for the planned migration-cleanup PR → leave as-is.
+
+So the only asset that genuinely needs a raster generator is the hero mascot.
+
+## Decision
+
+**Hybrid, not pure-AI** (this is the critic's decision-flipping correction):
+
+1. **Deterministic frame, AI-free.** Rebuild the hero's frame — background grid, layout, and
+ text chips — by **reusing the proven resvg SVG→PNG pipeline** already in
+ `prerender-seo.ts`. Text rendered from SVG = crisp, exact brand colors, **zero regen
+ drift**. AI never touches rendered text (dissolves the maskless-regen text-mangling risk on
+ short strings like "4 Packages").
+2. **AI only for the neon-cat mascot raster** — the one element that cannot be clean SVG
+ (soft neon bloom + gradients). Model: **`google/gemini-2-5-flash-image`** primary
+ (mascot has no text → Pro's text-rendering edge is moot; Flash is cheaper), escalate to
+ **`google/gemini-3-pro-image-preview`** (NanoBanana Pro) only if Flash quality is
+ insufficient. Both accept `images: [""]` (CLI auto-uploads — verified) +
+ natural-language edit prompt.
+3. **Two mascot variants** (keep-gold vs strip-to-blurple/pink) for operator/Maria to compare,
+ with cross-variant **consistency via reference-image** (feed variant-1 as `images[]` to
+ generate variant-2) — **NOT** a trained LoRA (LoRA training is overkill for two images).
+
+**Rejected candidates:** Grok (weak/unreliable text — but text is now SVG-side, so moot
+anyway); FLUX-Dev-LoRA (LoRA training overkill + text quality undocumented); Seedream/Reve
+(strong but Gemini dominates on instruction-edit + reference consistency). **Rejected
+approaches:** pure-AI whole-hero regen (maskless → text drift); full SVG rebuild of the mascot
+(neon-glow raster art does not translate to clean vector).
+
+## Consequences
+
+- **Positive:** text is deterministic + brand-exact; AI cost bounded to the mascot; no LoRA
+ training infra; reversible per-call (all models same CLI, pay-per-image, no lock-in).
+- **Negative / risk:** reference-image mascot consistency is **UNVERIFIED** — it is the single
+ load-bearing claim (critic's strongest objection). Must be pilot-tested before scaling.
+- **Neutral:** the neon cat becomes the documented primary mascot (currently undocumented).
+
+## Delivery channel (updated 2026-07-12)
+
+inference.sh CLI hit `insufficient balance` (zero credits). Rather than fund it, the operator
+has a **Gemini Pro (Advanced) subscription** which includes **Gemini 3 Pro Image / "Nano
+Banana Pro"** — the _same model_ as `google/gemini-3-pro-image-preview`. So the mascot
+generation runs via the **Gemini web app driven by Claude-in-Chrome**, at no incremental cost.
+The inference.sh CLI (`images: [""]` auto-upload, verified) remains the
+**scriptable fallback** once credits are added — preferred if the work ever needs to be
+batched/reproducible rather than one-off.
+
+## Pilot outcome (2026-07-12) — KEEP EXISTING MASCOT
+
+Pilot ran via Gemini web (Nano Banana Pro). **Technical result: PASS** — two consistent
+variants generated (gold-accent A; blurple-accent B), in-thread edit held pose/composition
+exactly (Gemini reported "Confirming Color Consistency"). So the critic's load-bearing
+consistency risk is **resolved: reference/in-thread consistency works.**
+
+**Aesthetic result: REJECT (operator).** Both regens read "too AI-ish" vs the original hero
+cat, which "feels more welcoming." Root cause: from-scratch regen launders out the original's
+asymmetry/hand-drawn character (the source of its warmth) into a generic uniform neon-sign
+look. This fires the **"pilot output worse than existing → keep existing"** trigger.
+
+**Decision: keep the original mascot as-is; do NOT replace it via AI regen.** The gold in the
+mascot contributes warmth the operator values — treat it as intentional, not drift. Any future
+palette tweak to the mascot must be an EDIT of the original raster (preserve its exact
+character), never a from-scratch regen — and even that is not currently wanted. AI image-gen
+via Gemini is a validated capability for OTHER assets (fresh graphics with no beloved
+predecessor), just not for re-deriving this mascot.
+
+## Iteration 2 (2026-07-12) — img2img RECOVERS the warmth
+
+Operator chose to keep iterating (better prompt + explore other skills) rather than abandon.
+Root cause of "too AI-ish" pinned by cropping the original: warmth = **simplicity + solid
+dark fill + chunky roundness + minimal detail** (sticker/kawaii), which from-scratch text
+regen launders into busy symmetric line-art. **Fix = img2img from the original crop** (not
+text-to-image): feeding the real cat as a reference preserved its exact cozy character while
+cleaning it up. Result accepted as the right direction.
+
+**Working method (Gemini web via Claude-in-Chrome, no paid CLI):**
+
+- Reference upload into the Gemini composer = **same-machine clipboard paste**:
+ `osascript … set the clipboard to (read POSIX file … as «class PNGf»)` then focus composer
+ - `cmd+v`. CRITICAL: the browser being driven and the machine running `osascript` must be
+ the SAME host — pasting the Mac clipboard into a REMOTE (Windows) Chrome types a literal "v"
+ (different clipboard). Use `list_connected_browsers` → `select_browser` to pin the local Mac
+ (`isLocal:true`) first.
+- `file_upload` MCP tool is BROKEN (rejects host paths, wants a `files` content param not in
+ its schema) — do not use; clipboard paste is the workaround.
+- Download-back is blocked: Brave prompts a native "Save As" dialog (unautomatable), JS
+ cross-origin fetch of the `googleusercontent` blob is CORS-blocked, curl 403s (needs
+ cookies). Operator downloads from their own browser, or fund inference.sh for scriptable
+ local-file img2img.
+- Prompt levers that kill the "AI-ish" feel: "SOLID dark-filled", "chunky/round", "minimal
+ detail, no busy line-work", "sticker/emote", keep exact reference.
+
+Skills installed to inform brand prompt craft: `anthropics/skills@brand-guidelines`,
+`inference-sh/skills@logo-design-guide`.
+
+## FINAL DECISION (2026-07-12) — KEEP THE ORIGINAL MASCOT
+
+After a `/debate` (5 lenses) and testing ALL three production paths, the operator chose to
+**keep the original mascot untouched.** Evidence trail (each shown to the operator, each
+rejected):
+
+1. **Text-to-image regen** (Nano Banana) → "too AI-ish" (busy line-art, generic).
+2. **img2img from the original** → preserved warmth but "too rounded, weird" (over-rounded blob).
+3. **Hand-authored vector SVG** (rendered via resvg — the production pipeline; neon-glow
+ filter DID survive resvg, so the vector path is technically viable) → "straight up ass":
+ thin strokes, weak glow, stiff geometric proportions vs the original's thick plump richly-
+ glowing warmth. Confirmed the debate's DESIGN-CRAFT prediction: vector-DIY hits a quality
+ ceiling for warm organic neon character.
+
+**Conclusion:** the original mascot is genuinely strong and none of AI-regen / img2img /
+vector-DIY beat it. Stop mascot work. If a definitive polished mascot is ever wanted, the
+reliable path is a **human illustrator** (debate plan B), not another generation attempt.
+**Redirect visual-identity effort to higher-ROI, taste-safe gaps:** the off-brand PURPLE
+`lucky-logo.svg` (deterministic recolor to Blurple/Pink) and missing product/dashboard
+screenshots in the README.
+
+## Logo close-out (2026-07-12) — RASTER SUFFICIENT, vector DEFERRED
+
+Second `/debate` (5 lenses) on "how to get a crisp scalable cat logo": **unanimous — ship
+the raster package, defer vector.** The raster (resampled from the 375px crop, gold kept)
+covers every use case for an 11-guild Discord bot (web logo, favicon, avatar, embeds). No
+print/large-format/merch need has been articulated, so crisp-vector ROI is contingent on a
+requirement that does not exist yet. Tracing the neon glow yields muddy geometry
+(pre-threshold mitigation = manual tuning = same fidelity ceiling the hand-vector hit).
+
+**Deferred vector fallbacks (only if a print/merch need materializes):** (a) bounded ~30-min
+image-to-svg trace of the line geometry + re-glow via SVG feGaussianBlur (render via resvg),
+show operator; (b) if that is muddy/stiff → commission a human illustrator. Vector/illustrator
+skills on skills.sh for that day: `shhac/skills@image-to-svg` (trace), `rknall/…@svg-logo-designer`,
+`sfkislev/flue@illustrator`.
+
+**Visual-identity thread CLOSED for now:** mascot = keep original; logo = the raster cat
+package (delivered, operator wires when ready); vector = deferred. Higher-ROI identity gap
+still open if wanted later: README product/dashboard screenshots.
+
+## Revisit when
+
+- **Pilot fails** (mascot inconsistent, wrong hue, or artifacts) → flip the mascot to manual
+ edit or keep the existing cat; do not iterate AI blindly.
+- **Gemini iteration exceeds 3 rounds** → stop, keep existing mascot or escalate; the cost has
+ overrun the SVG-alternative amortization point.
+- **A clean standalone mascot source** later exists (SVG/high-res raster) → reference-image
+ quality improves; re-pilot.
+- Text ever needs to go INTO an AI-generated asset → re-evaluate model text-rendering (Gemini
+ ~94% short-text, Reve typography engine) rather than defaulting to SVG.
diff --git a/decisions/2026-07-12-readme-presence-visual-proof.md b/decisions/2026-07-12-readme-presence-visual-proof.md
new file mode 100644
index 000000000..531e08277
--- /dev/null
+++ b/decisions/2026-07-12-readme-presence-visual-proof.md
@@ -0,0 +1,82 @@
+# README presence — ship public-landing visual proof; plumbing is trivial, not a blocker
+
+- **Date:** 2026-07-12
+- **Status:** Accepted
+- **Deciders:** Lucas Santana
+- **Method:** `/research-and-decide` — research agent (live-site + competitor survey) + artifact-only `decision-critic` (NEEDS_REVISION → revised after human-browser verification)
+
+## Context
+
+Lucky's original goal was "more presence." The GitHub README is text/table-strong (positioning,
+badges, feature/comparison/command tables, quickstart) but has **zero product visuals** — a wall
+of text. The mascot + logo threads already closed (see
+`decisions/2026-07-12-brand-asset-regen-tooling.md`); this decides where presence effort goes next.
+
+**Verified in a real browser (the critic demanded this; a fetcher had returned 403/404):**
+
+- `https://lucky.lucassantana.tech` is **UP** and **genuinely polished** — dark theme, neon-cat
+ logo, hero + CTAs, a "What Lucky does" 6-card feature grid, "A real web dashboard" section,
+ a dev/self-host architecture-card section, Apache-2.0 footer. Fully public (no auth).
+- The top.gg listing **exists** at bot id `962198089161134131` (the `839686019796811776` ref that
+ 404'd is a **wrong/stale bot id**).
+- Constraint: the _authed_ dashboard needs Discord OAuth login, which the assistant may NOT perform.
+ But the **public landing is assistant-screenshottable**, removing the operator bottleneck.
+
+## Decision
+
+**Ship visual proof from the polished PUBLIC LANDING, now, assistant-driven.** Concretely:
+
+1. Capture clean screenshots of the live landing (hero, "What Lucky does" grid, the
+ dashboard/feature section) and add them to the README with short captions. No auth, no operator
+ dependency, no design work.
+ **Delivered scope (this PR):** shipped `assets/screenshots/landing-features.png` (the
+ feature grid). The hero shot is deferred pending #1818 (landing still renders "Apache 2.0"
+ while the project is ISC) so the README doesn't advertise the wrong license.
+2. Fix the two verified plumbing bugs (trivial, not a blocker): the README live-demo link
+ (`href="https://lucassantana.tech"` → `https://lucky.lucassantana.tech`) and the wrong top.gg
+ bot id (`839686019796811776` → `962198089161134131`); add a top.gg badge/link.
+3. **Optional operator follow-up (not a blocker):** a real authed-dashboard screenshot or a short
+ Loom, if wanted — the landing already _claims_ "a real web dashboard"; a real shot would _prove_
+ it. Gated on operator (login), so it is a nice-to-have, done last.
+
+Framed honestly: this is **cheap, obvious polish** (empty README + polished landing already built),
+**not a validated growth lever.** No heavy measurement plan (YAGNI for a solo OSS bot).
+
+## Alternatives considered
+
+- **top.gg listing audit as the #1 lever** (research's top pick) — folded in as bug #2 (fix the id +
+ badge), but the listing already exists, so it is a small fix, not the headline.
+- **Bot-in-action demo GIF / screencast** — higher effort (needs a Discord test server + capture +
+ edit); deferred. A landing screenshot is far cheaper for equal first-impression lift.
+- **Authed dashboard screenshots as the primary asset** — rejected as the _primary_ because the auth
+ constraint makes the operator the bottleneck; demoted to optional follow-up.
+- **"Differentiate because competing music bots omit visuals"** — DROPPED. The critic correctly
+ flagged it as inferred from n=3 behavior, not evidence that visuals convert or don't.
+- **Architecture diagram / secondary bot-lists / mockup galleries** — low impact-per-effort, deferred.
+
+## Consequences
+
+- **Positive:** README gains real visual proof in one assistant-driven pass; two broken external
+ links fixed; the polished landing (already built, currently under-leveraged) starts working for
+ the project. Reversible (README edit).
+- **Negative / risk:** landing-UI changes will stale the screenshots (low maintenance cost — re-snap).
+ Presence may not be Lucky's actual growth bottleneck (discovery/community could matter more) — so
+ this is bounded polish, not a growth bet.
+- **Neutral:** authed-dashboard proof remains an open optional follow-up on the operator.
+
+## Revisit when
+
+- The landing page UI is redesigned → re-capture the README screenshots.
+- Evidence emerges that presence is NOT the growth bottleneck (e.g. traffic data shows the Discord
+ community drives adoption, not the README) → stop investing in README polish.
+- The operator wants to prove the _dashboard_ specifically → add the authed screenshot/Loom follow-up.
+
+## Related defects (found during this decision — file as issues)
+
+- README live-demo link href mismatch (bug #1 above).
+- Wrong top.gg bot id `839…` vs live `962…` (bug #2 above).
+- **License discrepancy:** README/LICENSE say **ISC**, the landing footer says **Apache 2.0** —
+ authoritative source must be reconciled (legal).
+- Node version drift: README says Node 22, the toolchain decision (`decisions/2026-07-12-*toolchain*`)
+ moved to Node 24.
+- Test-count claim conflict: README "~2500 tests" vs the social-preview hero "849+ Tests".