feat(eval): measure the context window, and name what still needs a model

[mrsibe]

Stacked on #197 (feat/eval-saturation-rule). Retarget to main after that merges.

What does this PR do?

Measures the window the prompt actually receives, and states plainly what still needs a model. Child 5 of #192 — the deterministic half.

The gap

The harness measured the retriever but never the context window. evidenceK: 5 was a separate constant from the contextK: 3 production uses, so the one metric that looked at a window looked at a different one than the product does.

The change

contextK is now the window for both context metrics:

• contextPrecision@contextK — of the first contextK passages, the share covering ground truth. This is what evidencePrecisionAt5 was, at the production width.
• contextRecall@contextK — the share of needed ground-truth blocks that made it into that window. Distinct from Recall@10: a block found at rank 4 is invisible when contextK = 3, and that is a product fact, not a ranking fact.

Both are deterministic: the dataset says which blocks answer the question, so no model is needed to score a window. Together they are the trade-off a contextK decision makes — a wider window finds more and carries more noise — which is what the sweep in the next child needs.

Baseline: contextPrecision@3 = 0.3556, contextRecall@3 = 1.0000. The needed evidence is always inside the top 3 on this corpus, but only about a third of what is inside the window is relevant. The second number is the one that says the window is paying for passages that do not answer the question.

evidenceK is removed from the harness config, so there is exactly one context width.

What is deliberately not here

Faithfulness, completeness, answer correctness and noise sensitivity are not added. They need a generative model, and the harness runs offline with only the pinned embedding model — the same constraint that keeps the reranker unmeasured (#170). This PR does not fake them. The report's Definitions section now says so explicitly, and the experiment script's "Not evaluated" note points at the same constraint.

Adding an LLM judge is a separate change that has to solve model pinning first. It is not a line of code, and shipping a judge that silently degrades to "no metrics" in CI would be worse than shipping none.

Testing

• npm run typecheck — clean
• npm test — 497 pass
• npm run eval twice — byte-identical docs/eval/baseline-v1.6.json
• npm run eval:retrieval — regenerated; hybrid still clears the amended rule

Related

Part of #192. Child 5 (deterministic half: context precision/recall).