fix(graph): sparse prop='fill' bit-parity with dense + upstream (changes metrics, needs review)

[frapercan]

Summary (CORRECTNESS-CRITICAL: changes metric numbers , please review)

The CAFAEVAL_SPARSE env switched not only the confusion-matrix kernel but also the propagation algorithm. Under prop="fill" (PROTEA's production default), the sparse path (_propagate_sparse_pushup) and the dense serial sweep produced different propagated prediction matrices on real DAG inputs, shifting Fmax:

| Setting | sparse (before) | dense (before) |
|, -|, -|, -|
| NK Fmax | 0.58859 | 0.58914 |
| LK Fmax | 0.61612 | 0.61391 |
| PK Fmax | 0.40195 | 0.40195 |

Which impl was canonical

The dense serial sweep is canonical: it is bit-identical to upstream cafaeval prop="fill" (BioComputingUP/CAFA-evaluator and claradepaolis/CAFA-evaluator-PK, the fork's ancestors). Verified on 5 hand-worked DAGs plus a 200-trial randomized stress; the dense path matched the upstream oracle in 100% of cases.

The sparse pushup was the buggy one. fill is a stepwise leaves-to-roots recurrence where a non-zero intermediate term blocks deeper descendants (it keeps its own value, and that value, not the descendant's, flows up to the parent). The pushup scattered every input non-zero straight into every ancestor and group-maxed, ignoring those blockers and overshooting. Worked counter-example (chain leaf->mid->root, leaf=1.0, mid=0.5, root=0):

• upstream / dense fill: root = 0.5 (mid blocks the leaf)
• sparse pushup: root = 1.0 (leaf scattered straight to root), WRONG

This matches the observed "sparse=1.0 where dense=0.86" divergence, concentrated at high tau on biological_process multi-term shared-ancestor rows.

The fix (which impl was fixed)

The sparse path is fixed. mode="fill" is now routed to a new _propagate_sparse_fill that reproduces the recurrence sparsely (walk ont.order, per-row max over each term's children's current values, never overwrite originally-non-zero cells). The fast scatter pushup is retained for mode="max", where it is proven correct. Minimal, localized change to the propagation path only.

After: bit-identical

• Propagated prediction matrices: bit-identical between sparse and dense, max|diff| = 0.000e+00 across BPO/CCO/MFO (including the BPO rows where the bug lived).
• NK/LK/PK Fmax now identical between both paths and equal to the dense (canonical) values:

| Setting | sparse (after) | dense (after) |
|, -|, -|, -|
| NK Fmax | 0.5891445 | 0.5891445 |
| LK Fmax | 0.6139093 | 0.6139093 |
| PK Fmax | 0.4019451 | 0.4019451 |

(Residual ~1e-16 / 1 ULP in some weighted count columns is float reduction-order noise in the counting kernel, downstream of the now-identical propagated matrices, below any metric significance.)

• Sparse stays ~5x faster than dense on the fill path (no perf regression).

Test added

tests/test_propagation_fill_parity.py: the worked blocked-chain counter-example, shared-ancestor and high-tau BPO-shaped rows, plus a 200-trial randomized DAG stress, all asserting sparse == dense == upstream-oracle bit-for-bit. Closes the coverage gap left by the existing self-parity test, which only exercises prop="max".

Local CI

ruff clean; mypy clean (py3.12 target, matching the CI runner); full tests/ + tests/diff (Phase B) green.

Deploy note

PROTEA pins cafaeval @ git+main. A redeploy will pick this up once it lands on main. I did NOT bump the pin.