night-shift: prompt-asset extraction (inline prompts -> versioned files + loader)

[syzayd]

What

Extracts the three inline system-prompt string constants (rag/pipeline.py's _SYSTEM,
agent/loop.py's _SYSTEM_TEMPLATE, review/weekly.py's _SYSTEM) out of Python source
into versioned plain-text assets under a new src/personal_llm/prompts/ package:
rag_system.txt, agent_system.txt, review_system.txt. Adds prompts/loader.py
exposing load_prompt(name) -> str, which the three call sites now use instead of
holding their own inline copy. agent_system.txt keeps its {tools} format placeholder
and doubled {{/}} JSON-example braces exactly as before - .format(tools=...) is
still applied by the caller, after loading, unchanged.

Also:

• pyproject.toml gained a [tool.setuptools.package-data] entry for
  personal_llm.prompts (*.txt) - without it, a real wheel build would silently drop
  the prompt files, since setuptools does not bundle non-.py files by default. The
  loader itself still reads via Path(__file__).parent (not importlib.resources) since
  that works identically in the source tree, an editable install, and a built wheel.
• eval/harness.py's docstring, which referenced the old _SYSTEM/_SYSTEM_TEMPLATE
  constant names, updated to match.
• README test-count badge/section: 118 -> 124.

graph/kg.py has its own separate inline _SYSTEM constant; left untouched, out of
scope for this task (only the three prompts named in the queue item were moved).

Why

PROJECT-GENESIS.md section 9, Tier 5, item #36 (aliased with Tier 9 item #78):
"personal-llm: prompt-asset extraction (inline prompts -> versioned files + tests)".
Verify clause: "tests confirm each prompt loads and round-trips."

How verified

• 6 new offline tests in tests/test_prompts.py: each loaded prompt compared
  byte-for-byte against a literal copy of the original inline string (not re-derived from
  the .txt file, so an accidental future edit is actually caught); unknown-name raises
  FileNotFoundError rather than returning empty; the {tools} placeholder round-trips
  through .format() with the doubled JSON-example braces surviving as literal single
  braces, not double-substituted.
• Deliberately corrupted agent_system.txt's doubled braces during development and
  confirmed the round-trip test failed loudly (KeyError from .format()) before
  restoring it - the test genuinely catches drift, not just agreeing with itself.
• Full suite: 124/124 passed (venv/bin/python -m pytest tests/ -q), 118 pre-existing +
  6 new, nothing weakened. The three existing call-site test files
  (test_rag.py/test_agent.py/test_review.py) pass unchanged, confirming the system
  prompt text sent to the (mocked) router is identical to before this change.