feat(zenpicker): i8 vs f16 agreement example + load_meta_picker_v0_1 + lib.rs additions

[lilith]

Summary

Single commit on top of the merged three-crate rename. Adds two zenpicker examples and lib.rs additions.

Files

• `zenanalyze/zenpicker/examples/i8_vs_f16_agreement.rs` — measures argmin agreement rate and max raw-output drift between i8-quantized and f16 bakes on the held-out 20% feature set
• `zenanalyze/zenpicker/examples/load_meta_picker_v0_1.rs` — loader example
• `zenanalyze/zenpicker/src/lib.rs` — 242 lines of additions

Test plan

• `cargo run --release -p zenpicker --example i8_vs_f16_agreement` against the validation TSV
• `cargo run --release -p zenpicker --example load_meta_picker_v0_1`
• Confirm lib.rs additions pass tests + clippy

[lilith]

Rebased onto current main (66d08f6) — NOT obsolete; the lib.rs additions are still novel.

Obsolescence assessment (per-piece, vs current origin/main zenpicker/src/lib.rs)

Piece	Verdict	Evidence on main
MetaPicker.family_at_output field + family_at_output() accessor	still-novel	main's MetaPicker struct has only predictor: Predictor<'b> — no cache field, no accessor
Subset-aware validate_family_order (parse + length check, subset OK)	still-novel	main still does exact-match if raw == ALL_LABELS_CSV { Ok } else { Err(FamilyOrderMismatch) }
parse_family_csv helper + 4 unit tests	still-novel	absent from main
Subset-aware pick (maps argmin idx through the bake's order)	still-novel	main's pick still does `pick.map(
pick_with_time_cost (size/time tradeoff via ArgminOffsets + ScoreTransform::Exp)	still-novel	absent from main
examples/i8_vs_f16_agreement.rs, examples/load_meta_picker_v0_1.rs	still-novel	zenpicker/examples/ does not exist on main

Nothing here is superseded — main's MetaPicker is still the pre-subset, exact-match version. So this is a "still applicable, needs adaptation" PR, not a close-candidate.

What changed during the rebase (916 commits behind)

Single API adaptation, driven by the zenpredict 0.2.0 change that made Model own its data and Predictor/MetaPicker borrow it:

• MetaPicker::new(model: Model<'b>) → MetaPicker::new(model: &'b Model). Conflict was confined to MetaPicker::new (this PR rewrote its docstring+body for the family_at_output cache; main independently changed the signature to &'b Model). Resolution: kept main's &'b Model borrow signature, folded in this PR's cache-on-construction logic (Predictor::new(model) and model.metadata()/.n_outputs() are both shared borrows, so no borrowck issue).
• Both examples updated to MetaPicker::new(&model) (Models bound before the pickers; they outlive the borrows).
• One clippy unusual_byte_groupings fix on the example RNG seed (0xC0FFEE_42DE_AD15 → 0x00C0_FFEE_42DE_AD15), surfaced now that --all-targets checks the examples.
• cargo fmt reflow of the new code to current rustfmt.

The bake feature is not referenced anywhere in this PR, so no bake→advanced migration was needed.

Green gate (local, on the rebased commit 3ed2ff4)

cargo build   -p zenpicker                          -> ok
cargo test    -p zenpicker --lib                     -> ok (10 passed; incl. 5 new: parse_family_csv_* x4 + time_cost_table_uses_codec_family_index)
cargo clippy  -p zenpicker --all-targets -- -D warnings -> ok
cargo build   -p zenpicker --examples                -> ok
cargo fmt     -p zenpicker --check                   -> ok

Force-pushed (lease) to feat/zenpicker-i8-agreement; tip is now 3ed2ff4d4d1b9db6c030ca651845055b5f1c4cd5, 1 commit on top of main (66d08f6), only the 3 in-scope files (+458/-19).

Note (not addressed, out of scope): i8_vs_f16_agreement.rs's header comment says it "falls back to a 1000-row synthetic sample if the TSV is absent" but the body always uses 5000 synthetic samples and never reads the TSV — pre-existing doc/code drift in the original commit. Both examples also point at absolute /home/lilith/oracle-d2-store/... bake paths, so cargo run needs those artifacts present; they build fine without them.

[lilith]

Closing without merging.

Decision: per-request family restriction is the caller's job via the existing AllowedFamilies mask, not the picker's. Bakes stay canonical — all 6 families in CodecFamily::ALL order, a fixed-shape function — and the caller restricts the choice set. The subset-bake machinery this PR adds (family_at_output, subset-aware validate_family_order / pick, parse_family_csv) buys non-canonical subset bakes, a flexibility we've decided not to take on. pick_with_time_cost is orthogonal/speculative (no time-aware meta-picker exists yet).

Branch feat/zenpicker-i8-agreement @ 3ed2ff4 (rebased onto 0.2.0, green) is preserved if any piece is revived — note the i8-vs-f16 agreement example there is a useful validation harness once its absolute /oracle-d2-store/... paths are fixed.