fix(consolidation): keep observations in their source facts' language

[nicoloboschi]

Fixes #3166.

Problem

build_consolidation_system_prompt() only carried a language rule when HINDSIGHT_API_LLM_OUTPUT_LANGUAGE was set — output_language_directive(None) returns "". Everything else in the consolidation prompt is English, so with no configured output language multilingual models drift and a bank of Chinese facts receives English observations. Retain has defaulted to preserving the input language since #181/#184; consolidation never did.

Reproduced against openai/gpt-oss-120b with the exact facts from the issue:

before → "The user often walks their pet in the park on weekends."
         "The user is preparing for a promotion defense."
after  → 用户周末经常带宠物去公园散步。
         用户最近在准备晋升答辩。

And on the update path — an existing English observation plus a new Chinese fact:

before → UPDATE: The user often takes their pet to the park for walks on weekends.
after  → UPDATE: 用户周末经常带猫豆豆去深圳湾公园散步。

HINDSIGHT_API_LLM_OUTPUT_LANGUAGE=English still forces English output from those same Chinese facts, and English facts still stay English (no over-correction).

Change

A ## LANGUAGE section is added to the consolidation system prompt only when no output language is configured. Beyond "keep the source language", it settles the three cases the issue asked to pin down:

• Per observation, not per batch — each observation follows the language of the facts in its own source_fact_ids. A mixed batch yields mixed observations; within one observation the majority language of its facts wins.
• Updates follow the new facts — an existing observation in a different language is rewritten entirely in the new facts' language, never left half-translated. Banks that already drifted converge back as new facts arrive.
• Names and technical terms are never translated — proper nouns, place/product names, identifiers, code, units stay verbatim.

When an output language is configured the new section is dropped rather than left to contradict "translate any source content into X".

Scope

This implements proposal 1 from the issue (prompt rule) and not proposals 2/3 (language detection + corrective retry + metrics), matching how retain solved the same class of drift. Two things worth knowing about that choice: the guarantee is prompt-level, which the docs now say explicitly; and _consolidate_batch_with_llm silently skips a batch once attempts are exhausted, so a "fail the batch" policy there would drop observations rather than surface an error. Happy to follow up with detection + retry if the prompt rule proves insufficient in the field.

Tests

• tests/test_multilingual_bm25.py — deterministic prompt-assembly coverage for both branches (rule present when unset, replaced by the directive when set).
• tests/test_consolidation_output_language.py (new, hs_llm_core) — real-LLM tests driving _consolidate_batch_with_llm and asserting via the LLM judge: Chinese creates, English observation updated by a Chinese fact, and the explicit-override case. Picked up automatically by the core-LLM CI job, which selects by marker.

Docs: hindsight-docs/docs/developer/multilingual.md documents the unset-language default, including the mixed-batch and update policies.