Integer donor-support snapping for discrete transfers; stacked-aware release gates

[MaxGhenis]

Fixes #794. Related: #782 (untouched — no bands or battery changes anywhere).

What

Three things, from the first release build ever attempted on a stacked pool:

1. Integer support at the transfer source. The 369 fractional weeks_unemployed values came from ACS-transfer QRF predictions on clone-1/2 rows (the post-transfer calibration hypothesis was tested against the real pool and refuted — that layer and its strict validator are untouched). Discrete-supported targets now snap deterministically to observed donor support (ties to the lower value), emit integer output, and the discrete-target roster is part of the transfer execution identity, so affected target-bank identities invalidate instead of silently changing content under an old receipt. Spec re-pinned coherently (42,122/42,122 fields, 41/41 inventory checks).
2. Stacked-aware weeks release gate. Source (LKWEEKS) reconciliation scoped to rows that have an ASEC source; channel roster derived from support-role metadata via a narrow read-only accessor (legacy frames keep working); the "5,218 PUF rows" misclassification fixed — those are ACS-origin nonnative clones. Every numeric threshold byte-identical.
3. Release-gate roster sweep. Clone-layout and physical-source assumptions modernized across the roster (SIPP Head Start/voluntary-filing clone handling, prior-year and SSI clone diagnostics generalized to all assembled clones, WIC draw identity made collision-free via person_source_id). Six archived-model cases (SSI disability criteria, SCF wealth, SCF auto loans, SIPP vehicles, SIPP tips, ORG wages/FLSA) need substantive ACS-predictor decisions and are documented with file:line evidence rather than patched with zero-fills or threshold changes — resolution in a follow-up PR.

A read-only replay against the real 25% pool after these fixes: the repaired checks pass; remaining failures are the 369 fractional weeks (fixed on rebuild), prior-year availability 0.0428 vs [0.05, 0.50], and is_pregnant=true on nonfemale rows — both genuine outcomes reported, not hidden.

Verification

Ruff clean; inventory verification green; all five pytest shards pass (build shard 6,608 tests, exit 0). Worktree clean. Root PROGRESS.md/out.md are the lane's session journals (history, not state).

Provenance: implemented by a gpt-5.6-sol lane; reviewed by Claude; no thresholds, bands, folds, seeds, or battery logic changed anywhere.

🤖 Generated with Claude Code

[MaxGhenis]

Superseded by #805 — this branch could never schedule CI (CONFLICTING vs #793 on main, so Actions had no merge ref to test). The gate-alignment work rides #805's integration branch (merged as 7bd7a04 with digests regenerated over the union tree).