feat: add passive prompt-cache controls

[Obed0101]

Problem

MendCode had no explicit, scoped control over passive provider prompt-cache behavior, and the ChatGPT OAuth adapter used one session-derived value for both transport affinity and cache identity.

Solution

• Add mendcode cache status|enable|disable with project, session, provider, and exact-model scopes.
• Preserve legacy provider behavior by default; enable smart cache keys only for verified API-key bindings and supported endpoints.
• Keep OAuth transport affinity separate from managed cache keys, rotate affinity when instructions change, and consume cache-disable markers locally.
• Add cache lineage fingerprints and conservative provider observations without keepalive, warm-up, or remote-cache deletion.
• Keep release contract tests on an isolated in-memory database in CI.

Verification

• Focused cache/OAuth tests: 210 passed, 0 failed (453 assertions).
• bun run typecheck: passed.
• Native candidate build and smoke test (bun run build --single --skip-embed-web-ui --skip-install): passed; --version and --help started successfully.
• Release-index tests: 3 passed.
• Release installation/channel/TUI smoke tests with isolated DB: all passed.
• git diff --check: passed.
• Full suite: 4616 passed, 12 skipped, 2 todo, 12 failed. The remaining failures are existing baseline/environment failures in session HTTP API, worktree shortcuts, SQLite identifier escaping, tool configuration, patch/freeform, snapshot race, and system-session coverage; this PR remains draft until required CI gates confirm the release path.

Risks and gaps

• Smart cache-key support is intentionally limited to verified OpenAI/OpenRouter API-key bindings; unsupported, custom, OAuth, and unknown bindings remain conservative.
• No live provider canary or cache-keeper behavior is introduced.
• The full-suite baseline failures remain disclosed rather than treated as release-ready evidence.