feat(ponytail): add minimal-code system-prompt injector toggle

[enjoyer-hub]

What

Adds a Ponytail mode next to Caveman: a toggle that injects a "lazy senior dev / write the minimum code that works" system prompt into outgoing requests, to curb agent over-engineering (fewer files, fewer needless deps, less boilerplate) — while never cutting input validation, error handling, security, or accessibility.

Caveman trims output verbosity; Ponytail trims over-building. They're complementary and can be enabled together.

Ruleset adapted from DietrichGebert/ponytail (MIT), the same way Caveman credits its upstream.

How

• open-sse/rtk/systemInject.js (new): the format-aware system-prompt injection (OpenAI / Claude / Gemini shapes) is extracted here from caveman.js, so both savers share one implementation (DRY).
• caveman.js: refactored to call the shared helper — behavior unchanged.
• open-sse/rtk/ponytailPrompts.js + ponytail.js (new): lite / full rulesets and injectPonytail.
• chatCore.js: injects ponytail just after caveman, gated on ponytailEnabled && ponytailLevel.
• src/sse/handlers/chat.js + settingsRepo.js: thread ponytailEnabled / ponytailLevel (defaults false / "full").
• Dashboard: a Ponytail toggle + level picker in the existing Token Saver card.

Tests

• tests/unit/ponytail.test.js — 10 cases: OpenAI/Claude/Gemini injection shapes, both levels, unknown-level + missing-body no-ops, and a guard that the shared refactor keeps Caveman working. All pass under the repo's vitest config.

 ✓ tests/unit/ponytail.test.js (10 tests)
 Test Files  1 passed (1)
      Tests  10 passed (10)

Note: pre-existing failures in tests/unit/rtk.test.js (it imports setRtkEnabled/isRtkEnabled, which open-sse/rtk/index.js doesn't export) are unrelated to this change — that file was not touched here.

Notes

• No new dependencies. Disabled by default; opt-in via the dashboard toggle.
• CLI was intentionally not changed, matching Caveman (which is also dashboard-only).