M4 disability gate: floors and anchor evidence base (#153)

* M4 disability gate: floors and anchor evidence base

Ceremony step 1 for the DI gate (roadmap #113, M4): the floors + anchor
basis its thresholds will derive from. Reads no gate, changes no gate;
gates.yaml is UNTOUCHED (the proposed gate_m4 stub-flip lives in the
artifact note + PR body for the referee rounds and eventual ratification).

ANCHOR-BASED (gate-1 external-check style), not holdout-only (gate-2): the
PSID EMPLOYMENT STATUS self-report is not the SSA DI program (seven named
concept deltas), so no SSA DI level is gated. Gated cells live only on
concept-bridgeable surfaces, scored against the sha-pinned in-repo DI ASR
2023 tables:
- (a) conversion-exit vs Table 50: retirement is the majority non-death
  disability exit near FRA (dominance, per sex).
- (b) prevalence age-shape vs Table 19: co-monotone rise + 50-59 peak
  (ordinal, per sex; bounded stats gated as invariants, not |ln ratio|).
- (d) internal incidence/recovery: the 100-seed person-disjoint half-split
  |ln(r_A/r_B)| floor, round(mean+3*sd,3) capped at ln(1.5).
Report-only with the delta named (never calibrated): the Table 49
termination-rate trend (period-pooling), the 6.B5.1 conversion level and
sex-ordering, all prevalence levels + the 60-FRA band, and the noisy
transition cells above the cap.

Deliverables: scripts/build_m4_gate_floors.py (deterministic; no populace-fit,
no pe-us), runs/m4_gate_floors_v1.json via write_new, binding tests (no dead
zones, label-swap catches, sha pins) + a PSID reproduction, tier refresh,
honest warts (k=4 empties the internal surface; the gate leans on the anchor
cells).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* M4 gate fixes round: referee findings 1-6 (comment 4950447491)

A/B (finding 1, BLOCKER): restore the pre-registered internal families
(incidence/recovery/prevalence/conversion) -- the concept deltas are PSID-vs-SSA
and never touch a candidate-vs-PSID reproduction cell. Adopt the referee's
computed mixed-k remedy: flow hazards @k=3 + prevalence.50-59 STOCK @k=4 (8
gated internal cells, faithful OC 0.9689). Commit the honest uniform-k
sensitivity over the pre-registered surface (k=2->21, k=3->12, k=4->2); correct
the conversion cells' machine reason to the power demotion.

C (finding 2, MAJOR): adopt the ratified 5200+k draw stream VERBATIM (was
4100+k) and carry the fresh_run_artifact_schema (per-draw rate cube [20,8,5],
undefined-draw invalidation, dispersion disclosure) into protocol + stub.

D (finding 3, MAJOR): pin the candidate-side anchor statistic -- per gate seed,
mean-over-K=20 rate basis, MARGIN >= MARGIN_K sigma (not the bare ordinal). Add
the degenerate table with the builder-computed epsilon-tilt row that FAILS the
pinned margin (margin_sigma ~0.09 << 3).

E/F (findings 4-6 + k rule): correct wart 2 (internal = level teeth incl. the
occupancy stock; anchors = exit-composition/shape teeth); name the T50
age-universe mismatch + mechanical dominance; label redundant anchor conditions
implied and frozen booleans corroboration; fix delta numbering (1-3 self-report,
4 transience, 5 conversion, 7 period pooling); name attrition censoring. Commit
the k-selection rule + the k=2 flow OC (0.2901) and the option-(i) uniform-k3 OC
(0.9008).

Full rebuild via the regenerator; updated always-runnable binding tests;
black/ruff clean; tier counts re-refreshed.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Author: MaxGhenis