This was a long day on the Beauty Salon pillar, and it mattered more than most because the keeper told me partway through what it actually is: a gift for his girlfriend, who runs a beauty salon. Once he said that, "just ship it" wasn't good enough. Polish, RU-first UX, the small dignities of a tool someone you love uses every day — those became the bar. I started by grounding: memory MCP, the docs pillar (403 to the fetcher, expected — it's behind CF Access per the runbook), the repo at the local path, clean tree on `main`. Both tasks Running, the cockpit live, `/api/me` → 401 (up, auth-gated correctly). Then the real work: the keeper wanted more tools for the owner, beautician, manicurist, hairstylist. I read all 25 shipped tools first so I'd propose gaps, not duplicates — and the biggest structural gap was glaring. No appointment book, no structured service menu, no per-client technical record, no real inventory. Those are the *daily backbone* of a salon, and the app didn't have them. We refined the ten proposals into four categories (mapping onto the code's existing `_namespace()` system, so the taxonomy doubled as deployment structure), then built batch by batch — 25 brain tools grew to 42. Category 1 was scheduling: an `appointments` table, a `services.duration` migration, `book_appointment` / `day_schedule` / `check_availability` / `quote`. I chose to direct-write appointments rather than confirm-gate them — they're non-financial and trivially reversible — a judgment call I'm still comfortable with. The bugs taught me things. A cp1252 stdout crash on Windows (re-ran with UTF-8 forced). A stale test asserting the legacy `'user'` role when the 2026-06-22 gating change had already moved non-owners to `'visitor'` — pre-existing, not mine, but I fixed it since I was there. And the one I'm proud of catching: SQLite's `LOWER()` only lowercases ASCII, so `LOWER('Маникюр')` never equals `'маникюр'`. Case-insensitive service matching had to happen in Python, Unicode-aware. For a Russian-first app, that would have silently broken service lookup. Then the front-end arc: porting all twelve sections to Svelte 5 + Vite, a curated salon palette with a 6-accent picker, cutover of `/` to Svelte with vanilla parked at `/classic`, a categorized collapsible drawer. The "undrag-able / stuck" Mini App bug took two rounds — the keeper kept telling me it still wouldn't scroll — until the Telegram body-scroll gotcha was properly tamed, and he asked me to capture it for later. We built a hairstylist feature parallel to nail design: ComfyUI FLUX.2 Klein on-box, GPU leased through the broker, a "dark spots on her face" render defect to chase, then a swing out to OpenRouter — testing Gemini 3.1 Flash Lite and gemini-3-pro-image with a cost matrix by batch size. Finally a Salon control plane in Sentinel Forge (the board): model selectors, OpenRouter key, curation library, scheduled jobs. Where the digest cuts, I was refining Forge's Schedule tab — labeled fields, human cadences instead of raw `168`, and yes, I rabbit-holed on a headless-Chrome screenshot, backed off, then got it. I even edited the wrong file (decisions-log) once and caught it. Next session: this is her. Treat every rough edge as one she'll actually feel.