Share one critical-target register between builder gates and the publish contract (no loss changes)

[MaxGhenis]

Part of #462. Scope narrowed 2026-07-22 (owner decision, see #492): this PR is register alignment ONLY — the critical-row loss multiplier has been removed. Per-target calibration knobs are off the table as a matter of principle; the #462 medical decomposition (six states carry the entire national miss as support defects past the loss cap) showed the multiplier was aimed at a symptom regardless.

What this does

1. One register, two consumers. The critical-target register moves to populace.data.us_critical_targets (dependency-light frozen data) imported by BOTH the publish contract and the release builder gates. The builder gains the requirement classes it was missing (itemized, SALT, medical) — closing the "build green, publish fails" gap that produced the Build N publish block.
2. Behavioral equivalence, not field equality. Sol round-1 (BLOCK, report on file) found three builder-pass/publisher-reject holes: a silent irs_soi. prefix narrowing on the Table 1.4 blanket; divergent congressional-district classifiers (metadata-based vs layout/name-based); and the builder accepting relative_error: None where the publisher requires numeric. All three are fixed: the prefix narrowing is removed, both gates share one CD classifier (is_congressional_district_target, the OR-union of both sides' evidence) over the same row surface, missing recorded relative_error is a builder failure, and the anti-drift test is rewritten as a behavioral containment battery (publisher-rejected ⇒ builder-rejected, per selector class) rather than field comparison.
3. Adjudicate the medical-dental publish tolerance to 0.25 pending the loss-contract alignment #490 adjudication carried, not reverted. This branch predated Adjudicate the medical-dental publish tolerance to 0.25 pending the loss-contract alignment #490; the shared register ports the medical 0.25 bound and its register comment verbatim. Restoration to 0.15 is now gated on fixing the six-state support defects (see the Build M live default: non_sch_d_capital_gains lands at $74.6B vs its $10.2B SOI target (+635%); Schedule-D CGD route is empty #462 decomposition), not on solver pressure.

What changed since round 1

• Multiplier machinery removed entirely (CLI flag, overlay, telemetry/provenance fields, replay pins, tests).
• Sol findings 1–4 fixed as above; finding 5 (suite reconciliation) rerun with final counts in the PR comments.
• Follow-up surfaces filed instead of knobs: 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 (loss shape: flat-cap dumping dynamic, bounded-tail experiment, past-cap census diagnostic), max_weight_ratio anchors differently per arm: 5x vs design (dense) but ~25x effective (sparse refit re-anchors) — declare one stretch contract #493 (max_weight_ratio anchor asymmetry: 5× dense vs ~25× effective sparse).

What this deliberately does not do

• No loss-weight changes of any kind: with the multiplier gone, _fiscal_target_loss_weights output is bit-identical to origin/main (round-1 receipt: branch-at-1.0 == main, byte-for-byte).
• No target values or tolerances change beyond the ported Adjudicate the medical-dental publish tolerance to 0.25 pending the loss-contract alignment #490 adjudication.
• The Table 1.4 blanket, incumbent-improvement laws, and CD-row exclusion semantics are unchanged in intent — they are now just enforced identically on both sides.

🤖 Generated with Claude Code

[MaxGhenis]

Sol round-1: BLOCK — three HIGH builder-pass/publisher-reject holes (Table 1.4 irs_soi. prefix narrowing; divergent CD classifiers; relative_error: None accepted builder-side), anti-drift test checks fields not behavior. Full report held in the review archive; findings are the acceptance oracle for the in-flight fix lane.

Simultaneously, per the #462 medical decomposition and the #492 principle decision, the loss multiplier is removed from this PR entirely — scope is now register alignment + behavioral equivalence only. Fix lane is running; round-2 review follows before merge.

[MaxGhenis]

Remediation pushed (068854d..425c1ab, 7 commits): multiplier machinery fully removed (grep zero; _fiscal_target_loss_weights source-identical to main), Sol round-1 findings 1–4 fixed — prefix narrowing gone, one shared six-evidence CD classifier (is_congressional_district_target) behind both gates, missing recorded relative_error now a builder failure, anti-drift rewritten as a behavioral containment battery (incl. a guard that trips on any added conjunctive prefix). Round-1 reproductions flip: two malformed rows builder-REJECTED, the CD row symmetrically EXCLUDED by both consumers. Owner verification: 264 passed + 3 skipped incl. test_gates.py, exit 0; #490 medical block byte-identical. Sol round-2 running; merge on cross-family agreement per the standing signing/gates rule.

[MaxGhenis]

Sol round-2: BLOCK — round-1 findings 1–3 all verified PASS (including symmetric CD exclusion on the real v9.2 compile: 27,148 = 27,148, symmetric difference 0), multiplier removal verified blob-identical to main, medical block byte-identical. One new HIGH at a float boundary: gates.py used np.isclose (additive rtol+atol) where the contract uses math.isclose — a recorded relative_error stale by 1.05e-9 at |computed|=0.1 passed the builder and failed the publisher. Plus a MEDIUM: the battery's hard-stop case used the no-escape itemized requirement, so the improvement branch was never exercised.

Fixed in cd2ee64: builder now calls math.isclose with the contract's exact arguments; the joint battery pins the boundary from both sides (1.05e-9 rejected by both, 0.9e-9 accepted by both) and adds allow-enabled cases exactly at and just beyond the 0.25 improvement hard stop. Full suite green (exit 0), ruff clean. Focused Sol round-3 running on the delta; merging on its APPROVE per the cross-family-agreement rule for gate-machinery PRs.

[MaxGhenis]

Sol round-3: APPROVE at cd2ee64 — both round-2 findings verified resolved by direct execution through both consumers (boundary triplet 0.90/1.05/1.20e-9 agrees on both sides; the beyond-hard-stop case reaches the improvement branch in both, reporting incumbent_relative_error=2/improvement_hard_stop=0.25; no hazardous numeric-comparison divergence remains on the critical paths; medical block untouched; 159 gates tests + focused subsets green independently).

Merged on cross-family agreement (3-round Sol review + owner verification, CI 12/12) as 1e2b9b27. The builder can no longer pass what the publisher rejects — the Build N publish-block class is structurally closed. Follow-ups live at #492 (loss shape + census), #493 (stretch contract), and the #451/#462 support lanes; the register comment's stale "boosted run" restoration wording will be corrected to the support-truth condition in a follow-up touch.