fix(split-phase1): generalize z-gap splitter for alphanumeric cell_ids

[atuldeshpande]

Summary

• _spatial_split_phase1_entities (the Split-Phase1 z-gap splitter in tests/_pipeline_runner.py) gated entities through a numeric-only regex ^\d+(-\d+){0,2}$. On datasets whose cell_ids contain dashes (PDAC, e.g. jikammne-1), the regex rejected every entity, so the entire z-gap stage was a silent no-op — entities that should split on a z-gap > dz_threshold (2.0 µm) were never examined.
• The gate now reads the _etype column (cell/partial) when present. _etype is set upstream from kernel codes (pruning.etype_from_codes), so it stays correct on alphanumeric cell_ids where label-string parsing is ambiguous. The legacy numeric regex is retained as a fallback when the column is absent, leaving that path byte-for-byte unchanged.
• The numeric collision pre-scan (next_suffix/next_subsuffix) and the m_main/m_part/m_sub mint branches collapse into a single prefix-agnostic counter that appends the next collision-free -{k} suffix to the parent label — behavior-equivalent on numeric ids, correct for any prefix.

Why

PDAC and other Xenium FFPE/IO datasets carry alphanumeric, dash-containing cell_ids. The numeric-only assumption silently disabled the z-gap splitter on exactly those datasets. This is the same root issue tracked under the entity-type-column refactor: infer identity/kind from the _etype column rather than parsing the label string.

Audit (other numeric-only label parsers)

• _phase1_rerank_within_parent_etype and _reassign_nuclear_post_1c_etype already anchor on the real cell_id column (lab.startswith(cid + "-")) and parse only the suffix — they handle adohnpem-1-style ids correctly. No change.
• The UNASSIGNED mirror _split_unassigned_components uses UNASSIGNED_{i} with a numeric connected-component index by construction — unaffected. No change.
• The dormant -tr- delimiter scheme (src/tracer/_etype.py) is the proper long-term fix but has zero live emitters; activating it requires regenerating all reference partitions, so it remains deferred to its own branch.

Tests

• New tests/test_split_phase1_zgap.py (6 direct unit tests): alphanumeric+_etype splits, numeric+_etype parity, numeric no-_etype legacy parity, alphanumeric no-_etype boundary no-op, no-gap control, and existing-child collision avoidance.
• Full suite: 157 passed, including test_pipeline_regression.py (reference-partition comparison) and test_etype.py — confirms numeric pipeline behavior is unchanged.

Test plan

• PYTHONPATH=src python -m pytest tests/test_split_phase1_zgap.py -x -q
• PYTHONPATH=src python -m pytest tests/test_pipeline_smoke.py -x -q
• PYTHONPATH=src python -m pytest tests/ -q (157 passed)

🤖 Generated with Claude Code