feat: reproducible train/eval split for post-training exports (#103)

[dcfocus]

Summary

Follow-up to #96: adds a group-disjoint, reproducible train/eval split to the export path, so eval scores aren't inflated by conversation/session leakage across the boundary. Closes #103.

Stacked on #111 (#96). Review/merge #111 first; until then this PR shows both commits (the split commit is last). The diff reduces to just #103 once #111 lands.

What it does

• SplitConfig { eval_fraction, by, seed } on ExportConfig. Each record's side is decided by a stable hash (FNV-1a + MurmurHash3 fmix64 finalizer) of its by grouping key + seed, so:
  • no group (default session_id; also run_id / tenant / source / bot_id / external-id prefix) spans both sides, and
  • the same seed reproduces the identical partition across runs/platforms (the hash is fully defined — not DefaultHasher).
• With split set, export_training writes <path>.train.<ext> and <path>.eval.<ext>, each with its own manifest recording the split params (eval_fraction, by, seed, side) and the complementary output path. The returned manifest is the train side.
• Curation (dedup/decontam/lifecycle from Curate lance-context records into trainable datasets (SFT/preference/rollout export) #96) runs once before the split, so the two compose: decontamination drops train rows near eval; the split keeps groups disjoint.

Python

ctx.export_training(
    "training/sft.jsonl",
    task="sft",
    split={"eval_fraction": 0.1, "by": "session_id", "seed": 42},
)
# -> training/sft.train.jsonl + training/sft.eval.jsonl (+ manifests)

Tests

• core (3): determinism (same seed → identical files), group-disjointness, fraction-within-tolerance, and split manifests with complement path.
• python (2): split disjointness + determinism.

README updated.

Checks

• cargo test -p lance-context-core --lib — 62 passed
• cargo fmt --all -- --check, cargo clippy --workspace --all-targets -- -D warnings — clean
• ruff format --check, ruff check, pyright — clean
• test_export_training.py — 10 passed

Acceptance criteria (#103)

• Split into train/eval by a grouping key with a fixed seed; re-running yields the identical partition
• No group (e.g. session_id) appears in both train and eval
• Each output carries a manifest recording split params
• Tests cover determinism, group-disjointness, and fraction accuracy within tolerance

Optional follow-up (noted in the issue): also emit the eval side in #98's labeled query-set format so retrieval-eval reuses the same artifact — deferred to keep this focused and avoid coupling to the (separate) #98 PR.

Closes #103