The per-chunk ontology list keeps to the prompt budget

[WaylandYang]

Closes #701.

Problem

With schema.org the extraction prompt was about 20k tokens per call. The budget in 0006 (ontology_prompt_budget, 24,000 characters) only decided whether the whole ontology fits. Once retrieval took over, the per-chunk list had no limit. Three floors added since then (ancestors, signature classes, relations declared on the retrieved classes) laid out 56,155 characters for one chunk: 2.3 times the budget, and 85% of the prompt. Descriptions were 84% of that list.

Change

• The per-chunk list keeps to the same budget. Candidates queue by distance: retrieved relations and attributes first, taking turns; floor additions after. Each candidate brings its signature classes. The list takes the longest prefix whose laid-out text (PromptLists::chars, the same measure the full list uses) fits the budget. Length only grows with the prefix, so a binary search finds it. Floor relations still pull no signature classes, as before.
• A retrieved list carries each description's first sentence. Sentence bounds follow UAX Fix GLIBC mismatch, and make the release prove the container boots #29 through unicode-segmentation (already in the tree). A full inline list is a small ontology and keeps descriptions whole.
• 0006 status line, a revision entry and the README row are updated.

Measured

Recall bench (NVIDIA filings, 52 hand-checked items, schema.org), measured together with #705 on the same run:

	before	after
recall	47/52	48/52 (1σ ≈ 2.7)
prompt per extraction call, same 76 chunks	13.5k	8.1k
total prompt tokens, same 76 chunks	1,029,899	617,215

The heavier case, google-deepmind.txt from ai-timeline (41 chunks, re-extracted with both PRs): prompt per call 19,473 → 8,300 tokens, output 1,605 → 965.

Tests

• a_description_is_cut_at_its_first_sentence, two_ranked_lists_take_turns_and_the_longer_one_finishes
• cargo test -p utopia-extract -p utopia-server, clippy and fmt clean

🤖 Generated with Claude Code