The Feno program - Fenocosm record v35: the corrected truth-referenced recovery readout (P-9.7) - v32's N*_recovery withdrawn as a model-misspecification artifact

Fenocosm record v35 — the corrected truth-referenced recovery readout (P-9.7), beside record v32. A sealed, append-only results record of the Feno program, and a correction of one of the program's own headline negative results. Record v32 split the event-altitude estimability floor into two constants and reported that one of them, N*_recovery, is above the grid at every amplitude and on both observation processes — that on data carrying a genuine k = 6, S = 40 hierarchy the fitted multiplier runs +45 % to +86 % high with a bias that never shrinks with N, and that 61 of 72 sealed cascade track-rows return a span on its own parameter infimum. On that reading the program retroactively demoted every fitted ladder span it had published from a measurement to a model-selection quantity. That reading is a model-misspecification artifact, and this record withdraws it. The sealed floor grids build the cascade world on exposure × activity_window — a hard-gated raised cosine that at the pinned duty 0.30 is exactly zero in ~70 % of bins and peaks near 6.7× — but fit every cascade entrant with covariate = exposure alone. That omission is correct for the selection race (the periodic-window entrant is precisely the rival being raced, so handing the window to the cascade would decide the race by construction) and wrong for a truth-referenced recovery readout: it spends the six-layer ladder absorbing a deterministic 16-bin comb instead of the injected hierarchy. This record refits the same sealed cells, from their own pinned seeds, at the true depth k = 6, with the generator's own covariate and the fitter settings otherwise unchanged (n_starts = 3, maxiter = 250, kappa fixed at 1). On the canonical track: v32 (covariate-free)v35 (generator's covariate) N*_recovery, multiplierabove the grid100 multiplier recovered at N = 1,6000 / 6 seeds6 / 6 seeds median m̂_hi at N = 1,600 (injected 1.8)2.959 (+64 %)1.791 (−0.5 %) median Ŝ at N = 1,600 (injected 40)2.33646.7 multiplier bias over the re-analysed N+51 % … +79 %−4 % … +2 % track-rows with median S on the infimum18 of 192 of 19 The corrected multiplier-recovery floor is 100 (v20 trivial exposure), 200 (v28 real-anchored CHIME transit) and 140 (v28 trivial) — an order of magnitude below N*_selection ≈ 800, and deep inside the sealed grid. The readout also refits six cells with v32's own covariate-free likelihood as a byte-level control and reproduces the sealed per-cell fits to 3.6 × 10−7, so the streams are the sealed streams and the covariate is the only thing that changed. What stands, confirmed rather than retracted. The split itself — that model selection and parameter recovery are different floors — is confirmed; this record re-measures one side of it. N*_selection ≈ 800 is untouched (the covariate-free fit is the right design for the race, and the selection column is read from the sealed rows), and it now carries the uncertainty it never had: a six-seed panel crossing a ≥ 5/6 rule on a strongly non-monotone curve, Clopper–Pearson 95 % on 5/6 = [0.36, 1.00], grid bracket N* ∈ [400, 1,600]. The ladder-degeneracy flag (S ≤ 1.05 or m_hi ≥ 104 ⇒ boundary solution) and both diagnostic repairs are unaffected. And the span is still not recovered: with the corrected covariate it stops collapsing and its medians land in the right decade (46.7 against an injected 40), but the per-seed factor-2 gate is still not met by ≥ 5/6 seeds at any N in range (best 4/6). So "a fitted span is a selection quantity until shown recovered at the actual N" survives as a rule; what does not survive is the claim that the multiplier is unrecoverable at event altitude. Scope, declared before the run and sealed with the record. The canonical amplitude track of each sealed grid plus every track a claim surface quotes: v20 | trivial (N = 50…1,600), v28 | transit and v28 | trivial (N = 50…800) — 120 refits, 20,349 CPU-seconds. The weak/strong amplitude brackets, the v20 | sessions track and the larger N values were not re-run, and their v32 rows stand as read under the covariate-free likelihood; the record says so rather than implying grid-wide coverage. The N truncations are a declared compute budget, re-declared once on a measured per-cell cost and recorded as such: the reduction removes the most expensive cells — evidence dropped, not evidence selected — and no aggregate had been computed when it was made. One cell of the first scope completed before the change and is reported as a single-seed, out-of-scope corroboration rather than suppressed. The apparatus is strictly additive: no sealed constant and no sealed code path is touched, so records v0–v34 regenerate byte-identically. Reproduction requires no external data — every stream is re-simulated from a pinned seed. Deposit note. All seven Feno-program repositories are private and remain private, so this deposit is standalone: it carries its own reproduction materials, its metadata contains no repository link, and sibling work is referenced by DOI rather than by repository path. Related Zenodo DOIs cited below are reserved with publication pending. Part of the Feno program. The source repository is private; the reproduction materials for this record are contained in the record itself.