[rohan-hotdata]

This PR makes the vector ranking the default of HotdataStore. It also lets capture store a session as chunks, so the reader gets the parts of a session that match. Closes #26.

How

• Default ranking. HotdataStore(ranking="vector") is the default now, as MemoryStore ranks. "fused" stays as an option. The bench runner defaults follow.
• Chunks. Memory.capture(..., episode=key, chunk_chars=n) cuts the text at blank lines outside fences into chunks of at most n characters and stores chunk i under <key>-<i>. It does not use the heading rules of load, because an assistant turn often holds a Markdown heading. The extractor reads the whole text once, and each fact cites every chunk. A retry writes nothing, and an edited chunk gets its next revision only. No new Store method or record field.
• Block display. A fact that cites 11 chunks showed 11 ids in its block line. The block now shows a run of adjacent ids <prefix>-<n>@<rev> (same prefix, n counting up by 1, same revision) as <prefix>-<first>…<last>@<rev>. Other sources show as before. The stored sources do not change. A rule that folded by key prefix alone was rejected: a caller's own keys chat-0002 and chat-0003 look like chunks.
• Bench. bench.retrieval --ablate ranks with the fused query less one arm. bench.e2e --chunk-chars stores chunks, and the chunk size is part of the kept name. bench.diagnose maps a chunk key to its session.
• Skill. scripts/capture.py --chunk-chars.

Why

The ablation on the 419 retrieval questions showed that the BM25 arm makes the whole gap between the rankings: fused recall_all@10 0.9570, vector 0.9928, BM25 alone 0.8926. Episodes have no cues, so fused is BM25 plus the content vector. Fusion with equal weights pushes evidence sessions that BM25 ranks 23 to 40 out of the top 10. End to end on sample A at 64,000 characters, vector scored 46 and fused 45.

A session is one episode, so one needed turn brought the whole session into the block. With chunks of 1,000 characters at a budget of 16,000:

Sample	Chunks of 1,000, 16,000	Phase 8, whole sessions, 64,000
A	48 of 54, 4,336 reader tokens	45, 12,849 tokens
B	46 of 54, 4,360 tokens	45, 12,676 tokens

The full grid (1,000 / 2,000 / 4,000 at 8,000 / 16,000 / 32,000) and the diagnosis are in docs/benchmarks.md.

Verification

• make verify and make integration (144 hotdata tests) on the final library commit.
• New tests cover chunked capture on both drivers, retries, edits, refusals, and the range display. Others prove the same block from MemoryStore and HotdataStore with chunks, for "search" and "cited". Two more prove that capture searches with the store's ranking, and that the default is vector. The skill test runs the new --chunk-chars command.
• Benchmarks: every number in docs/benchmarks.md comes from a result file in bench/results/. Phase 10 spent $12.77 of the $50 cap.
• A docs audit of every shipped doc against the final diff.

Notes

• Behavior change. Callers of HotdataStore without ranking now get the vector ranking. The block format also changes for runs of chunk sources. Both are in CHANGELOG.md.
• "cited" is unchanged. A fact cites every chunk of its session, so "cited" still gives the whole session. Only "search" saves tokens with chunks.
• Extraction with chunks is not identical. A fact line with a range is a little longer. The known-fact list that the extractor reads holds 2,000 characters, so it can hold one fact less. Part of one keep run paid for extraction again.
• At 32,000 with chunks of 1,000, the block held 29.9 chunks, so k=30 and not the budget was the limit.
• BM25 favors long sessions a little (length percentile 0.645 against 0.583 for vector). This was raised at the task 1 stop, and judged a cost of the arm, not of the data.
• Found on the way: hotdata-framework 0.14.1 list_managed_databases reads only the first 100 databases, and HotdataStore.provision uses it. Filed as list_managed_databases and resolve_managed_database read only the first page of databases sdk-python-framework#89. Not fixed here.
• The roadmap closes phase 10 in a separate PR after this merges.