Benchmark assign_compartments transport pruning and gap-filling

[edkerk]

Two assign_compartments features shipped with correctness tests but no performance measurement — flagged as the exposed flank when reviewing what the compartment-assignment work can do. This adds a reproducible ablation benchmark and a study doc measuring both on yeast-GEM (flattened to one compartment and reassigned).

Transport pruning (prune_transports) — a clean win

	transports added	reaction agreement	runtime
pruning on (default)	994	0.721	209 s
pruning off	1268	0.721	74 s

Pruning removes 274 transports (21.6 %) — a strict subset of the unpruned set — at identical accuracy (0.721 to four decimals). It only ever removes genuinely redundant shuttles; it never trades a placement. The cost is runtime (re-certifies after each removal), not quality.

Gap-filling — precise and conservative

It never fires gratuitously. On the natural flattened draft, gap-fill adds 0 reactions with or without a universal — transport addition alone restores growth. The genome-scale confirmation of the test_no_gratuitous_gapfill unit test.

When it fires under real gaps, it adds the right reaction. Ground-truthed knockout-recovery (remove each growth-essential reaction, gap-fill from a universal containing it — the same cobra.flux_analysis.gapfill call _gapfill wraps):

outcome	of 60
recovered, re-added the exact removed reaction	45 %
cobra numerical failure (declined)	55 %
recovered with a wrong reaction	0 %

• 100 % of recoveries are exact; zero wrong additions. When gap-fill answers, it answers correctly.
• The only failure mode is declining to answer. The 55 % shortfall is entirely cobra.flux_analysis.gapfill's numerical-validation limit, which _gapfill catches — so an unsolvable gap stays uncertified and visible, never silently mis-filled.
• cobra's error suggests lowering integer_threshold; a sweep shows its default is a sharp optimum (every other value, up or down, collapses recovery to 0 %), so there is no tuning win and _gapfill correctly uses the default.

Verdict

Transport pruning is a strict win; gap-filling is precise, conservative, and fails safe. Both are now measured, not just asserted.

Notes

• Docs- and script-only; no source change. pytest 827 passed / 4 skipped; ruff and the -W docs build clean. The knockout sample is seeded (deterministic).
• Study: docs/studies/assignment_ablations.md; reproduce with python scripts/benchmark_assignment_ablations.py.