feat(eval): LLM-backed case generation (ir.eval_gen)

[thorwhalen]

What

Adds ir.eval_gen — the build-time case generator that completes the eval harness (#12). It turns a corpus into DiscoveryCase data by back-translation (description → user intents; the artifact id is the free gold label), to be scored offline by ir.eval (PR #18).

This is PR 2 of the eval harness: scoring landed in #18 (ir 0.1.3); this is generation.

Design

• Name masking (anti-leakage). The artifact name is scrubbed from the description before generation, and any intent that still reuses the name is dropped. Masking is whitespace/separator-tolerant (ci-advisor / ci advisor / ci_advisor / ciadvisor), and the output guard is bag-of-words aligned with vd's BM25 — reordered multi-word name tokens count as a leak, a single shared content word does not. This keeps the lexical leg of hybrid from trivially matching query→gold.
• Abstention slice. A fraction of cases are "no artifact applies" (empty gold), sized to hit a target fraction.
• Injectable LLM. query_generator / abstention_generator are callables, so all generation logic is hermetically testable with stubs; the oa-backed defaults are lazy-imported (so import ir.eval_gen stays cheap/offline).
• corpus_signature() anchors a frozen case file to its (live, machine-specific) corpus.
• CLI: ir eval-gen <corpus> <out.jsonl> (needs oa; scoring stays offline).

Review & testing

The diff went through a multi-agent adversarial review (correctness · masking robustness · oa-fidelity · test gaps → verify). All confirmed findings fixed, notably:

• _parse_lines used lstrip(charset), which corrupted real queries (3D→D, -9 degrees→degrees, .env→env) — now strips only genuine list markers.
• Multi-word masking required adjacency, so reordered name tokens slipped the guard while BM25 still matched — now whitespace-tolerant + token-set guard.
• Name coercion for malformed name fields; deterministic max_artifacts (sorted); input validation (k>=1, abstention_frac ∈ [0,1)).

33 hermetic eval-gen tests (94 total green); ruff clean.

Real end-to-end smoke

Generated real cases from the live skills corpus and scored them — the question that motivated "eval first":

mode	NDCG@10	recall@1	recall@5
dense (MiniLM)	0.845	0.625	1.000
hybrid (MiniLM+BM25+RRF)	0.694	0.375	0.875

On name-masked, paraphrastic discovery queries, dense beats hybrid — masking strips the rare identifiers the BM25 leg needs, so its signal becomes noise. (Preliminary: n=8 gold; a full generated set is the real measurement.)

Refs #12, #1.