Keogh carriers lack selection support in the frozen rmloss100 57k: protect-swap for Build M + restoration-time support checks

[MaxGhenis]

Build M sparse attempt 11 failed the keogh_distribution_neutralization reform-coverage smoke probe with a structural zero (+$11.8k vs the ≥$1M floor; expected ~+$24.47M). The gate is working exactly as the coverage campaign designed: it caught a restored input that the certified selection cannot express.

Diagnosis (measured on base-m, sha 5d8a3162):

• keogh_distributions is CORRECT in the base pool: $148.97M weighted across exactly 13 carrier households (13 persons, all positive) — matching the probe's own expectation from the locked ASEC DST_SC*/DST_VAL* source (Populate remaining US tax-input and reported-observation layers #38).
• The rmloss100 selection (Build I, 152baca3) predates the keogh restoration, so its 57,240 households include only 4 of the 13 carriers — the tiny ones (~$0.1M unweighted). An input restored after the selection froze has no selection support; no count-based gate can see it (the column is zero on the incumbent reference too, so parity passes trivially). Only the smoke probe has teeth here.

Remedy shipped with Build M: buildm-keogh-carrier-protect-swap — a documented protect-swap in the rmloss100 lineage pattern (rmloss100 itself was a 100-household misc swap):

• ADD the 9 absent carrier households.
• REMOVE the 9 lowest-base-weight non-carrier asec-channel households (deterministic tiebreak on the join key; base weights 22.7–27.4, the thinnest records in the selection). Channel composition preserved; n stays 57,240.
• Result: the full $148.97M weighted source mass is in-selection under base weights; the calibration re-solve re-equilibrates the 9 removals (0.016% of records).
• Manifest: buildm_keogh_swap_selection_source.json, identities sha256 1812b819b0a2…, parent 152baca3…, constructed against base-m 5d8a3162….

Systemic follow-up for architecture week: every future input restoration needs a selection-support check at restoration time (carriers ∩ frozen selection, with a protect-swap or reselection decision recorded), so this class is caught before a release run instead of by the smoke gate two builds later.

[MaxGhenis]

Shipped in the certified Build M: the protect-swap selection manifest (identities sha 1812b819…) carried all 13 keogh households through selection, and the certified run's smoke scored keogh at +$118.1M (vs the structural-zero failure this issue diagnosed). The systemic follow-up also landed: #436's preflight check_smoke_probe_support performs exactly the restoration-time carriers-∩-selection check this issue prescribed (it caught the pre-swap gap statically), and #445/#446 added the mass-protection layer the swap design turned out to need. Certified verdict: populace#368 comment (2026-07-17).