refactor(llm): move the production prompts into files behind one loader

[jasonssdev]

Summary

First slice of #1277: the 13 production system prompts and 3 prompt fragments move out of Python string constants into one folder, one file each, behind one loader. No prompt wording changes — every rendered prompt is byte-identical to before.

• ADR-0043 (Proposed): LLM prompts are files versioned by content hash.
• Layout: src/openkos/prompts/<task>/<name>.md, 16 files in 11 task folders (adjudication, contradiction, edge_typing, volatility_typing, reconciliation, decision_subject, decision_revision, answer, extraction, rationale).
• Loader: src/openkos/llm/prompts.py, stdlib-only, reads through importlib.resources like templates/, as bytes decoded UTF-8 with no stripping (several prompts end without a newline, one with a blank line). .gitattributes marks the folder -text so a checkout cannot rewrite line endings.
• Placeholders: four prompts splice in values the code also uses (extraction/system, edge_typing/system, answer/system, answer/sufficiency); their files hold {{name}} placeholders that load_prompt fills. The loader refuses an unfilled placeholder or an unused argument.
• Compatibility: every module-level constant keeps its name (e.g. contradiction._SYSTEM_PROMPT), now loaded from its file, so evals that import or .replace() them are unaffected. SUBJECT_PROMPT_VERSION and JUDGE_PROMPT_VERSION are derived through prompt_hash with identical values, so the revision-findings cache is not invalidated.
• docs/architecture.md tree updated.

Not in this PR: stamping prompt identity into eval results (point 5 of the issue).

Related issue

Refs #1277

Type of change

• feat — new feature
• fix — bug fix
• docs — documentation only
• refactor — no behavior change
• test — tests only
• chore / ci — tooling, build, or CI
• Breaking change

How was this tested?

• Byte-identity oracle (tests/unit/test_prompt_byte_identity.py, its own commit): records the rendered hash of 16 prompt constants plus the 2 cache-key versions. It passed on the old code and passes unchanged after the move — 16/16 and 2/2 equal.
• Registry pin (tests/unit/llm/test_prompts.py): one table of prompt id → file hash; fails on any edited or unregistered file, and covers the loader's refusals and the no-carriage-return rule.
• Mutation: appending one byte to contradiction/system.md fails the pin and the oracle.
• Wheel: uv build ships all 16 .md files; installed into a clean venv outside the repo, the loaded prompts hash the same as the oracle. tests/smoke/test_cross_platform_wheel.py gains test_every_prompt_file_ships_and_loads.
• ruff check, ruff format --check, mypy ., pytest --cov (96.46%), evals/run_self_tests.py (46/46) pass locally.

Checklist

• My commits follow Conventional Commits.
• I added or updated tests for the change.
• I updated docs where behavior, interfaces, or the knowledge model changed.
• Lint, format, type check, and tests pass locally (ruff, mypy, pytest).
• Output remains OKF-conformant and derived stores stay reconstructible from the bundle + sources.
• The change is consistent with the project's guiding principles (local-first, provenance, freshness, human-in-the-loop).