[rohan-hotdata]

This PR adds a test of cues on facts to the benchmark, and closes phase 11. The configuration does not change.

How

• bench/prompts.py: the extractor variant detail-cues. It is detail, plus one to three questions that each fact answers, returned in a schema with cues. Variant gains cues and a schema property.
• bench/e2e.py: the extractor puts the cues of the reply on each Fact. --recall-ranking recalls with another ranking than the one that captured, so one kept memory is recalled with both rankings. The result file records it as memory.recall_ranking.
• bench/diagnose.py: decides the cause of a miss with the recall ranking of the run.
• docs/benchmarks.md: the section "Cues on facts", and the variant table. bench/README.md: both options.

No change to src/hotmemory.

Why

Until now the benchmark extractor wrote no cues, so the cue arm of every fused run was empty. The open question of phase 11 was whether cues make the fused ranking better than the vector ranking on facts.

Memory	Recall	Sample A	Sample B
detail	vector	31	33
detail	fused	26	32
detail-cues	vector	28	32
detail-cues	fused	30	34

Cues lift the fused ranking (+4, +2), but no arm beats detail with the vector ranking on both samples. The cue prompt writes about 22 percent fewer facts. The diagnosis of sample B found 0 of 20 and 1 of 22 misses ranked below k, so the ranking has little to gain on these small stores.

Verification

• make verify: ruff, mypy, 307 tests passed. New tests cover the cue variant, --recall-ranking in Config, and that run_question recalls through the recall-ranking store.
• The eight runs and two diagnoses above, on the local stack, $6.94 in total. The detail capture replayed from the cache at $0.
• Not run: make integration, because no driver changed.

Notes

• A store where the answer competes with thousands of facts, such as a BEAM conversation, can give cues and the fused ranking more to gain. Not measured.
• recall uses one ranking for facts and chunks, and chunks have no cues, so the production configuration with the fused ranking was not tried.