i18n: honor source-content language for extraction labels

[motuo520]

Problem

_EXTRACTION_SYSTEM (graphify/llm.py) is entirely English and code-oriented, and
nothing in it tells the model what language to write labels in. LLMs follow the
system prompt's language, so non-English corpora get English labels.

Reproduced with a 4-document Chinese corpus (deepseek-v4-flash / qwen via
OpenAI-compatible endpoint), graphifyy 0.9.16:

• Before: 28 nodes, 0/28 Chinese labels (e.g. Chinese notes on Kantian ethics
  produced "Kantian Deontology", "Categorical Imperative", ...). Hyperedge labels
  likewise English.
• After appending the language rule below: 27/27 Chinese labels, and edges
  even improved (17 → 21).

Proposal

Append a language rule to the extraction system prompt. Two variants; either works.

Minimal (always-on, recommended) — models are good at detecting the dominant
language, no config surface needed:

LANGUAGE RULE: Write all human-readable text (node labels, hyperedge labels) in
the dominant language of the source content. Chinese source content MUST produce
Chinese labels with verbatim Chinese concepts (e.g. 道德义务论, 绝对命令) — do NOT
translate them to English. Node IDs stay ASCII per the format rules above.

Configurable — GRAPHIFY_LABEL_LANGUAGE=auto|zh-CN|en|..., default auto
(= the rule above); an explicit value pins the label language regardless of
source language (useful for bilingual corpora where the user wants uniform labels).

Notes

• Node IDs intentionally stay ASCII ([a-z0-9_]); this change only affects
  human-readable labels.
• Existing graphs keep their labels until rebuilt — no migration concern.
• Happy to open the PR with tests if you tell me which variant you prefer.

Suggested patch (minimal variant)

--- a/graphify/llm.py
+++ b/graphify/llm.py
@@ _EXTRACTION_SYSTEM tail @@
 ...],"input_tokens":0,"output_tokens":0}
+
+LANGUAGE RULE: Write all human-readable text (node labels, hyperedge labels) in
+the dominant language of the source content. Chinese source content MUST produce
+Chinese labels with verbatim Chinese concepts (e.g. 道德义务论, 绝对命令) — do NOT
+translate them to English. Node IDs stay ASCII per the format rules above.
 """

[graphify-labs]

Graphify reviewed this change.

Looks safe to merge — no coupling regressions and no blocking issues, checked against the code graph (not a self-assessment).

---

Graphify review — findings

Adds a language rule to the semantic extraction prompt in both graphify/skill.md and skills/graphify/skill.md, instructing labels to use the source content's dominant language (Chinese concepts stay verbatim, not translated) while node IDs remain ASCII.

No blocking issues surfaced.

Analysis details — impact, health, verification

Impact & health

Graphify review

Impact — 60 functions depend on the 60 functions this change touches.

Health — grade A; no new coupling hotspots.

Verification — 60 functions in the blast radius were not formally verified this run (proofs are advisory here).

Health delta baseline: last indexed commit b14b52e (diverged from this PR's base — delta is approximate).

Gate & verification

graphify gate

PASS — objectively clean (no health regressions, tests not run — proofs not run this pass (advisory)). Grounded, not self-assessed.

Advisory (not blocking):

• verification_scope: 60 function(s) in the blast radius were not formally verified this run