feat(sepia): prose-layer rows for Opus 4.8, Opus 5, Fable 5, and Gemini 3; record consulted absences

Follow-up to #45, which introduced the vendor-guidance prose layer with
two tables (Claude Fable 5.1, GPT-5.6). This fills in every other vendor
page that says something about its own model's writing, and records the
searches that found nothing so they are not repeated.

Claude gains three version-tagged prose tables in model-fingerprints.md:
Fable 5 / Mythos 5 (un-steered elaboration — option surveys, root-cause
essays, heavily structured PR descriptions, narrating comments; and the
long-session arrow-chain shorthand), with the vendor's brevity
instruction embedded verbatim and operative for that release as the
executor; Opus 5 (longer default responses that effort does not shorten,
padded written deliverables, ready narration of intent and of
corrections), with the vendor's conciseness instruction embedded
verbatim; Opus 4.8 (direct, opinionated, minimal validation-forward
phrasing, sparing emoji; length calibrated to task complexity), where
the handling column notes that the absence of validation openers is the
release default, not evidence of a human. Gemini gains a series-tagged
table from the Gemini 3 developer guide (less verbose by default,
direct and efficient, persona only when prompted); SKILL.md's version
definition gains one clause so a release inside a vendor-scoped series
matches that series' table.

Absences are recorded as facts with dates rather than left as "not yet":
Sonnet 5's page says only that prose style may shift; Opus 4.7, Opus 4.6,
and Sonnet 4.6 have no model-specific page and are covered only by
generation-level statements in the best-practices reference; DeepSeek's
API documentation has no prompting guide and its V3.2 model card gives
sampling parameters only; Kimi's "Best Practices for Prompts" is generic
with no version named and the K2.5 model card's default system prompt
was removed 2026-01-29. Those releases and families load nothing for the
prose layer, and research/sources.md gains a "Consulted with no
prose-layer statement" table alongside five new vendor-guidance rows.

Verification: the three embedded verbatim blocks and the four quoted
fragments in table cells were compared against the live vendor pages on
2026-09-03 and matched (whitespace-normalized substring check); the
Gemini sentence likewise. The 26 measured narrative rows are unchanged
from main. Unit suite 41 green; check_versions unchanged at 0.4.1.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TtoUEQarAYtQYrZLLwqjao