refactor(PR2): safe deletions — remove 14 dead / out-of-scope symbols

[nceglia]

Stacked on #30 (PR0–PR1). Base = refactor/pr0-contract-freeze — GitHub retargets to main when #30 merges.

Pure removal PR: deletes the Phase-2 dead / out-of-scope symbols outright (dropped, not moved to examples). No behavior change to the kept surface; import tcri green; full suite 35 passed / 1 skipped.

Deleted — 14 symbols, 384 lines

• preprocessing: get_latent_embedding, group_small_clones, register_probability_columns, remove_meaningless_genes, gene_entropy, classify_phenotypes
• plotting: polar_plot (undefined-name ref), probability_distribution (self-recursion bug), bayesian_mutual_information (bad kwarg), SankeyNode.hex_to_rgb
• metrics: _ent (never wired in), clone_fraction, module-level dkl (replaced by _distance.kl_divergence in PR1)
• utils: probabilities (+ its dead plotting import)

Safety

Every target verified to have zero in-package call-sites before removal — including a precise NAME( call-site check for the three ambiguous ones (_ent, module-level dkl, probabilities — all 0). The surviving "dkl" references are the string distance-metric name + the flux inner dkl_func (kept → Phase 6), not the deleted function. Removal done by AST-span (exact def line ranges), not text munging.

Audit

A 3-lens multi-agent audit (doc↔code · deletion safety · plan/contract conformance) ran on this branch → PASS, with 3 LOW items, all fixed here:

• dropped the cosine_similarity import this PR orphaned (sole user was the deleted classify_phenotypes);
• corrected a diary line-count (−129 → −127);
• struck a stale plan line that still marked classify_phenotypes as a Phase-4 fold (it's a Phase-2 DROP per the authoritative REDO_LIST.md).

The audit independently confirmed: all 14 on-plan Phase-2/DROP, none present in the frozen _contract.pyi kept surface, no under-deletion, import + suite green.

Ledger: docs/contract/REFACTOR_AGENDA.md Phase-2 block fully ticked; PR-table row 2 → ✅.

🤖 Generated with Claude Code