refactor: unify model residency control, gate the OpenAI UI, and remove phase jargon

[quiet-node]

Overview

Addresses the three Phase 2 follow-ups tracked in #220, all pre-release cleanups deliberately kept out of the stacked feature PRs (#217/#218/#219) to keep those diffs focused. The feature is unreleased, so the config-shape change here is migration-free.

What changed

• Unified model-residency control. The built-in engine and Ollama now share one setting, keep_warm_inactivity_minutes, with one consistent meaning: -1 keeps the model resident forever, 0 uses the provider's natural short default (about 5 minutes), and N unloads after N idle minutes. The redundant idle_unload_minutes field is removed. A small tested translator maps the field onto the built-in runner's internal timer; the Ollama keep-alive path is unchanged. Settings now shows a single "Keep Warm" control for both local providers (hidden for a remote OpenAI-compatible server, whose residency Thuki does not manage). The built-in engine's effective default moves from "resident forever" to "unload after about 5 minutes idle", which frees several GB of RAM when idle now that built-in is the default provider on every fresh install; power users can set -1.
• OpenAI provider UI gated behind a compile-time dev flag. Thuki ships local-only (built-in plus local or remote Ollama). The Settings affordances to create or manage an OpenAI-compatible provider are now behind VITE_ENABLE_OPENAI_PROVIDER, off by default and tree-shaken out of shipped builds. The backend stays fully active and tested (the /v1 client shared with the built-in engine, Keychain storage, the routing arm, the config kind); this gates only the user-facing kind, so no end user can reach an OpenAI provider in a shipped build.
• Internal phase jargon removed. "Phase 1/2/3" and "Phase A/B" references in comments, section headers, and test labels are reworded to describe behavior. The screenshot algorithm's "Phase 1/2" local steps become "Step 1/2". The animation phase state and the mmproj download phase are left as-is (real domain terms).

How it works

The unified field is read at the two boundaries that already forward residency policy: startup engine seeding and the Settings change-forwarder. Both translate through one warmup::builtin_idle_minutes helper (-1 to the runner's "disabled, stays resident forever", 0 to a fixed about-5-minute timer, N to N minutes), so the runner's internal "0 = disabled" convention stays an implementation detail behind one tested boundary. The OpenAI gate reads import.meta.env.VITE_ENABLE_OPENAI_PROVIDER inline at the one render site that mounts the OpenAI card and its add-a-server affordance; Vite statically replaces the value at build, so a production build folds it to false and tree-shakes the affordance out of the bundle.

Testing

bun run test:all:coverage and bun run validate-build both clean: frontend at 100% across lines, branches, functions, and statements; backend 100% line coverage; lint, format, typecheck, and build all clean. The residency translator has unit tests for every arm; the OpenAI flag is tested in both enabled and disabled states.

Closes #220.