Rowwise local solve surface under the doctrine (#495 increment 6b)

[MaxGhenis]

What

Increment 6b of #495: the solve shape the rowwise candidate uses — one weight vector over the cloned households, where each household supports only its assigned constituency's target rows (the US Build-N structure).

• build_uk_rowwise_local_matrix: one column per household; an area's rows draw support only from households assigned there; fails closed when the target surface misses an assigned area (the 650/650 requirement — misses are support or target work, never silent exclusion); reuses the stacked path's align_area_targets contract.
• solve_uk_rowwise_weights_under_doctrine: structurally knob-free exactly like the UK local solve doctrine: uniform operator, declared bounds, past-cap census (#495 increment 4) #503 stacked doctrine solve (no per-target parameters, no doctrine injection point, duplicate-surface refusal), with the one rowwise-correct difference: initial weights are the household base weights floored at min_initial_weight — never split across areas, because a rowwise household exists in exactly one area.
• Shared core extracted as solve_prepared_local_weights: the stacked path routes through it unchanged (pinned by a regression test asserting the split-initialization behavior), and both shapes carry the Loss shape: flat cap at 1.0 makes past-cap rows free dumping grounds (17 rows pushed out in Build N); bounded-tail experiment + past-cap census diagnostic #492 past-cap census.
• rowwise_area_support_summary: per-area assigned/nonzero households, distinct source households, weight mass, ESS — over every target area, zeros included.

Testing

8 new tests (matrix placement asserted cell-by-cell against hand values, uncovered-area/alignment/finiteness refusals, base-weight initialization pinned, knob-freedom signature + TypeError, duplicate-surface refusal, support summary hand-checked, stacked-behavior regression) plus the existing doctrine/solver/runner suites green (36 total in the solve family). Ruff clean.

Part of #495.

🤖 Generated with Claude Code

[MaxGhenis]

Cross-family review completed (sol adversarial pass, fable adjudication) — all three blockers confirmed and fixed in the second commit:

1. Unreachable target rows (nonzero target, zero support — sol reproduced error 1.0 with zero past-cap counts): the builder now fails closed with named (area, metric) examples; zero targets with zero support stay legal.
2. Solver→writer mass fence: rowwise_calibration_mass_record constructs the record a calibration must append, and the solve docstring names the contract — with the precision that the Rowwise clone: carry weight-kind/mass-log fences, exact mass conservation, pool lineage, dry-run plan (#495 increment 2) #501 writer's chain-currency fence already refuses an unrecorded mass change at write time (sol's mass-drift repro would fail the write), so the gap was the missing record constructor and documentation, not an unfenced write path.
3. Dead-row resurrection: zero base weights are refused outright — the floor must not revive IMPORTANCE-zero rows; reviving is upstream work with a recorded mass change. min_initial_weight=0 now allowed (stacked parity, finding 10).

Should-fixes all fixed: prepared-core validation (NaN/shape/alignment/empty-surface, finding 4), metadata-name collision refusal (finding 5), named duplicate-label refusal (finding 6), support-summary input normalization incl. composite ids (finding 7), restored stacked error fidelity (finding 8), and the nine tests (finding 9). Sol's bitwise-equivalence verification of the refactor across ordinary/conserving/L0/budget solves (finding 11) is the strongest evidence the shared core preserved behavior.

41 solve-family tests green; ruff clean.