feat(prompts): register author-published prompt variants (#20, #87) (#204)

* feat(prompts): add verbatim author prompt-variant data module

Byte-exact copies of the alternate prompt templates each judge's authors publish
(Prometheus grading templates, Selene YAML templates, GLIDER prompt + data
wrappers, ShieldGemma prompt-only/prompt+response, PolyGuard full taxonomy,
DynaGuard full reasoning format). Harvested from raw author sources and stored
json-encoded so intentional author whitespace (e.g. trailing spaces) survives the
trailing-whitespace hook. Stdlib-only leaf, imported by prompt_registry.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(prompts): register author-published prompt variants (reference-only)

Add each judge's author-published alternate prompts as additional named versions
alongside the runtime default: Prometheus (6 grading templates + 2 systems),
Selene (5), GLIDER (canonical + 3 data wrappers), ShieldGemma (prompt-only,
prompt+response), PolyGuard (full S1-S14 system), DynaGuard (full reasoning
format). Each is overridable=False with a verified author source URL — for
discovery, copy/adaptation, and #194 pinning, not drop-in runtime swaps (their
placeholder/output contracts differ from the guardrail's default pipeline).

resolve_prompt now rejects a reference-only version selected via prompt_version=
with a clear ValueError (previously it would fail with a cryptic KeyError deep in
_pre_processing); get_prompt/list_prompt_versions still expose them for discovery.

The harvest also revealed three defaults were mislabeled provenance=author when
their text actually diverges from the author source (Prometheus adds a 'Feedback:'
prefix, PolyGuard drops the whole S1-S14 taxonomy, DynaGuard simplifies the
answer format), so their labels are corrected to 'adapted' and the true verbatim
author text is registered alongside. Byte-exactness re-verified against a fresh
fetch of every cited source; provenance independently audited (all authoritative).

Regenerates schemas/guardrail_prompts.json and docs/prompt_catalog.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(prompts): cover reference-only enforcement + verbatim author quirks

Assert every non-default variant is reference-only + author-provenance + sourced;
that resolve_prompt rejects reference-only versions while get_prompt still returns
them and inline prompt= is honored; that the data module is an import-free leaf;
and that author quirks are preserved byte-for-byte (GLIDER 'must in' typo + trailing
space, ShieldGemma double space, Prometheus's no-'Feedback:' output line, PolyGuard's
embedded S1-S14 taxonomy).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* docs(prompts): document author-published prompt variants

Correct the guide's claim that variants are swappable via prompt_version= (they are
reference-only), add a Prometheus example listing/inspecting variants, and link the
browsable prompt catalog. Note that selecting a reference-only version at runtime
raises ValueError.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(prompts): address Copilot review on #204

- resolve_prompt's reference-only error now suggests
  AnyGuardrail.get_prompt(GuardrailName.<NAME>, ...) (enum form) instead of a
  bare string, so the copy-pasted snippet type-checks.
- Reword the _authored_prompt_data docstring to describe the actual encoding
  (single-line escaped literals; no whitespace at a physical line end) rather
  than 'double-quoted' (ruff-format switches some literals to single quotes to
  minimize escapes).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: dni138