[{"type":"thinking","thinking":"The user says proceed with the next stage: repairs 4.5a–4.5d. Order per STATE.md: 4.5a first (D-036 §1 reference-point rewrite + D-039 repair + I_I housekeeping), then 4.5b (symbolic concavity), then 4.5c (valley lower bounds — hard), 4.5d (bookkeeping write-ups).\n\nThis is substantial mathematical work. Let me plan:\n\n**4.5a tasks:**\n1. D-036 §1 rewrite: prove |R(s)| ≤ A log(|t|+2) + B₀ with traced constants using the reference-point trick.\n   Classical argument (Davenport Ch. 15): for s = σ + it, −3 ≤ σ ≤ 4 (our range), |t| ≥ 10:\n   ζ′/ζ(s) − Σ_{|γ−t|<1} 1/(s−ρ) = O(log t).\n   \n   Reference-point approach: start from partial fractions for ξ′/ξ:\n   ξ′/ξ(s) = B + Σ_ρ (1/(s−ρ) + 1/ρ).\n   Then ζ′/ζ(s) = ξ′/ξ(s) − 1/s − 1/(s−1) + (1/2)log π − (1/2)ψ(s/2)... wait actually ξ(s) = ½s(s−1)π^{−s/2}Γ(s/2)ζ(s), so ξ′/ξ = 1/s + 1/(s−1) − ½log π + ½ψ(s/2) + ζ′/ζ. Thus ζ′/ζ = ξ′/ξ − 1/s − 1/(s−1) + ½log π − ½ψ(s/2). (matches D-036 line 16).\n   \n   R(s) := ζ′/ζ(s) − Σ_{|γ−t|<1} 1/(s−ρ)\n         = B + Σ_{|γ−t|<1} 1/ρ + Σ_{|γ−t|≥1} [1/(s−ρ) + 1/ρ] − 1/s − 1/(s−1) + ½log π − ½ψ(s/2).\n   \n   The problematic piece: Σ_{|γ−t|≥1} s/(ρ(s−ρ)). The reference-point trick: compare with the same expression at s₀ = 2 + it... hmm, but wait: Davenport's standard lemma is for −1 ≤ σ ≤ 2; our range extends to σ = −3 (actually −5/2 in the contour? D-036 uses σ₀ = −5/2 vertical) and σ up to 4 or 2√n... hmm, I_III is the LEFT vertical at σ₀ = −5/2, length 2T. So we need the bound for σ ∈ [−5/2, 2√n]... the R-bound is used for the vertical segments and horizontal ones. Range: −5/2 ≤ σ ≤ 2√n, |t| ≥ 10 (plus short piece M₀ separately).\n\n   Hmm wait — actually for σ ≥ 3, |ζ′/ζ| is directly bounded by |ζ′/ζ(2)|·2^{−(σ−2)} (Dirichlet series) — no partial fractions needed. The R-bound is needed for −5/2 ≤ σ ≤ 3.\n\n   Reference-point trick: write the difference\n   ξ′/ξ(s) − ξ′/ξ(s₀) where s₀ = 2 + it (or 3+it):\n   = Σ_ρ [1/(s−ρ) − 1/(s₀−ρ)].\n   For |γ−t| ≥ 1: 1/(s−ρ) − 1/(s₀−ρ) = (s₀−s)/((s−ρ)(s₀−ρ)), |s₀−s| ≤ |σ−2|... bounded by 4.5 (σ ∈ [−5/2, 2])... for σ beyond 2 use direct Dirichlet series. And |s−ρ| ≥ |t−γ| ≥ 1 (since |β−σ| ≤ ... hmm: |s−ρ|² = (σ−β)² + (t−γ)² ≥ (t−γ)² ≥ 1. And |s₀−ρ| = |2+it−ρ| ≥ ... ≥ |t−γ| but more: |s₀−ρ| ≥ max(|2−β|, |t−γ|) and also |s₀−ρ| ≥ c|ρ| for... need Σ 1/(|s−ρ||s₀−ρ|) ≤ Σ 1/(t−γ)² — but careful near γ where (t−γ) small excluded. Σ_{|γ−t|≥1} 1/(t−γ)² ≤ 2C₀(t)(1 + π²/6·...) — as in Lemma D. So far-sum difference ≤ 4.5·2C₀(t)·(1+π²/6) = O(log t). \n   Near sum: Σ_{|γ−t|<1} [1/(s−ρ) − 1/(s₀−ρ)] — the 1/(s−ρ) terms stay in R(s)'s definition (subtracted), the 1/(s₀−ρ) terms: |s₀−ρ| ≥ ... for |γ−t| < 1, |2−β| ≥ 1 (β ≤ 1) so |s₀−ρ| ≥ 1... Actually |s₀ − ρ| ≥ |Re| = 2−β ≥ 1. So Σ_{|γ−t|<1} 1/|s₀−ρ| ≤ C₀(t)·1. Good.\n   So: ζ′/ζ(s) − Σ_{|γ−t|<1} 1/(s−ρ) = [ζ′/ζ(s₀)] + [ξ′/ξ difference adjustments]... let me do this cleanly:\n   \n   ζ′/ζ(s) = ξ′/ξ(s) − 1/s − 1/(s−1) + ½log π − ½ψ(s/2).\n   ξ′/ξ(s) = ξ′/ξ(s₀) + Σ_ρ[1/(s−ρ) − 1/(s₀−ρ)].\n   ξ′/ξ(s₀) = ζ′/ζ(s₀) + 1/s₀ + 1/(s₀−1) − ½log π + ½ψ(s₀/2).\n   \n   Assemble:\n   R(s) = ζ′/ζ(s) − Σ_{|γ−t|<1}1/(s−ρ)\n        = ζ′/ζ(s₀) + [1/s₀ − 1/s] + [1/(s₀−1) − 1/(s−1)] − ½[ψ(s₀/2) − ψ(s/2)] + Σ_{|γ−t|<1}[−1/(s₀−ρ)] + Σ_{|γ−t|≥1}[1/(s−ρ) − 1/(s₀−ρ)].\n   \n   Bounds:\n   - |ζ′/ζ(s₀)| ≤ |ζ′/ζ(2)| = 0.5699 (s₀ = 2+it: |ζ′/ζ(2+it)| ≤ −ζ′/ζ(2) = 0.5699 — certified: compute to certified digits; the D-054 F5 item: get a certified bound, e.g. via flint/interval or just bound by Σ Λ(n)/n² = 0.5699... — the Dirichlet series Σ Λ(n)n^{−2} converges absolutely; bound the tail rigorously: Σ_{n≤N} + tail ≤ (integral bound). Easy certified 0.57.)\n   - |1/s₀ − 1/s| ≤ |s−s₀|/(|s||s₀|) ≤ 4.5/(|t|·|t|) small.\n   - |1/(s₀−1) − 1/(s−1)| ≤ 4.5/(|t−1|²)-ish small.\n   - ψ difference: ½|ψ(s₀/2) − ψ(s/2)|: ψ(w) = log w − 1/(2w) + r(w) — THIS is where the digamma remainder (D-039) is needed, for w = s/2 with Re w ∈ [−5/4, 2], |Im w| ≥ 5. Difference: ψ(s₀/2) − ψ(s/2) = log(s₀/2) − log(s/2) − [1/(2s₀) − 1/(2s)]·... + r(s₀/2) − r(s/2). |log(s₀/2) − log(s/2)| = |log(s₀/s)| = |log(1 + (s₀−s)/s)| ≤ log(1 + 4.5/|s|)·... ≤ 4.5/|s|·(1/(1−4.5/|s|)) small. And |r| ≤ 1/|w|² = 4/|s|². So the ψ-difference ≤ O(1/|t|) — GOOD, no log term needed from ψ! Wait, but then where does log t come from? Hmm: ψ(s/2) individually is log(s/2)-sized, but the DIFFERENCE is small. But wait — the difference approach gives R(s) = O(log t)?? Let me recount: the pieces are ζ′/ζ(s₀) (bounded), 1/s terms (small), ψ-difference (small), near-sum at s₀ (≤ C₀(t) ~ 0.43 log t), far-sum difference (≤ 4.5·Σ1/(t−γ)² ~ C₀ π²/6 ~ 0.7 log t). Total: R(s) ≤ ~1.2 log t + O(1). There it is — the log t comes from the ZERO DENSITY C₀(t), not from ψ. \n   \n   Wait, hmm: but is that right? The classical bound is ζ′/ζ(s) = Σ_{|γ−t|<1} 1/(s−ρ) + O(log t). Yes. And our far-sum difference estimate: Σ_{|γ−t|≥1} |s₀−s|/(|s−ρ||s₀−ρ|) ≤ 4.5 Σ 1/(t−γ)² — with Σ_{|γ−t|≥1}1/(t−γ)² ≤ 2C₀(t)·π²/6 (unit-interval grouping; zeros in [t±k] ≤ C₀, 1/(t−γ)² ≤ 1/k²) — ~ 0.72 log t + const. Plus near-at-s₀ sum ≤ C₀(t)·1 (|s₀−ρ| ≥ 1 — actually ≥ 1 since Re(s₀−ρ) = 2−β ≥ 1... |s₀−ρ| ≥ 1 yes).\n   \n   So traced: |R(s)| ≤ 0.57 + small + [C₀(t) + 4.5·2C₀(t)π²/6] + ½·(ψ-diff ~ 9/|t| + 8/|s|²)... \n   C₀(t) = 0.433 log(t+2) + 2.5. So |R| ≤ C₀(t)[1 + 4.5·2·π²/6·] wait: 4.5·2C₀·(π²/6)?? Let me redo: Σ_{k≥1} 1/k² over both sides: Σ_{|γ−t|≥1} 1/(t−γ)² ≤ 2·C₀(t)·(π²/6)? Hmm — unit intervals [t+k, t+k+1) and [t−k−1, t−k): zeros ≤ C₀(t+k) ≤ C₀(2t)-ish... C₀ monotone: C₀(t+k) ≤ C₀(2t) for k ≤ t. For k > t (zeros far above): C₀(t+k) ≤ C₀(2k)... need care. Standard: C₀(t+k)/k² — split k ≤ t and k > t. For k ≤ t: C₀(t+k) ≤ C₀(2t) = 0.433 log(2t+2)+2.5 ≤ 0.433 log(t+2) + 0.433·log 2 + ... ≈ C₀(t) + 0.31. For k > t: Σ_{k>t} C₀(t+k)/k² ≤ Σ (0.433 log(t+k+2)+2.5)/k² ≤ (0.433 log(2t+2)+2.5)·(1/t + ...)... small O(log t/t). Fine — constants assembled in the write-up.\n   \n   So the reference-point rewrite gives |R(s)| ≤ A log(|t|+2) + B₀ with A ≈ 0.433·(1 + 4.5·2·π²/6)?? Hmm wait: near-sum at s₀: Σ_{|γ−t|<1} 1/|s₀−ρ| ≤ C₀(t)·max 1/|s₀−ρ|. |s₀−ρ| ≥ 2−β ≥ 1 ⟹ ≤ C₀(t). Hmm but actually could be up to 2 zeros per unit interval... C₀(t) is the unit-interval count bound, fine.\n   Far: 4.5·Σ_{|γ−t|≥1}1/(t−γ)² ≤ 4.5·[2(C₀(t)+0.31)·π²/6 + O(log t/t)] ≈ 4.5·1.645·C₀ ≈ 7.4 C₀. Hmm that's big: total ≈ (1+7.4)C₀ = 8.4·(0.433 log + 2.5) = 3.6 log(t+2) + 21. That's A = 3.6, B₀ = 21 vs measured A = 0.9, B₀ = 1.0. The traced constants are lossy (factor 4 in A, 21 in B₀) — but still O(log t)! The question: does R-004 survive A = 3.6, B₀ = 21? B(n) leading coefficient 0.577 + 2A = 0.577 + 7.2 = 7.8 √n log n vs S_∞ ½ n log n. Crossover: ½n log n > 7.8√n log n ⟺ √n > 15.6 ⟺ n > 243 — leading-order fine! Lower-order terms: B₀ enters I_III as 4T·B₀ = 84√n and H(T) = 3C₀² + A log + B₀ in horizontals ~ 2(e⁴+1)·7·(3C₀² + 3.6 log + 21) ~ 760·(3·(0.433log+2.5)² + ...) — at n = 5e4: log n ~ 10.8, C₀(√n) = 0.433·log(225)+2.5 ≈ 0.433·5.4+2.5 = 4.84; 3C₀² = 70; H ≈ 70 + 3.6·5.4 + 21 = 110; 2(e⁴+1)·7·110 ≈ 780·110 ≈ 86 000?? vs margin at N₀ = 5e4 currently 67 705 with the tighter constants. Hmm — with A = 3.6, B₀ = 21 the margin at 5e4 might go NEGATIVE, pushing N₀ up. N₀ can move up — the finite verification covers to 5e5 (Misguich–Voros), so N₀ anywhere below 5e5 works. Need to recompute the crossover with the new constants: B(n) ≈ 7.8√n log n + ~2(e⁴+1)·7·(3C₀²+...) + 84√n + 2N(√n)... at n = 5e5: √n = 707, log n = 13.1: S_∞ ≈ ½·5e5·13.1 = 3.3e6; B ≈ 7.8·707·13.1 = 72 000 + horizontals 780·(3·(0.433·6.56+2.5)² + 3.6·6.56+21) ≈ 780·(3·28.5+44.6) ≈ 780·130 = 101 000 + 84·707 = 59 000 + 2N(707) ≈ 2·707/2π·log(707/2πe)·2... small ~ 2400. Total B ≈ 235 000 < 3.3e6 — margin at 5e5 hugely positive. Crossover somewhere ~n = 1e4–5e4. Fine — the theorem survives with N₀ possibly moved up (still ≪ 5e5, 10× overlap preserved or reduced). E-023 rerun with new constants will determine the exact N₀. \n\n   So 4.5a plan: (i) write the reference-point derivation with traced constants (new D-056 or edit D-036 §1 — better: new derivation D-056 that supersedes D-036 §1, keeping D-036 as historical record with its erratum); (ii) repair D-039 digamma bound (needed where? In the reference-point version, ψ-difference needs |r(w)| ≤ 1/|w|² for Re w ≥ −5/4, |Im w| ≥ 5 — same lemma); (iii) I_I housekeeping (contour length 2(√n+1), certified ζ′/ζ(2)); (iv) rerun E-023 with traced constants to re-derive N₀/margins — E-023b? No, E-023b is the acb box cover. Call it E-028 (margin rerun with fully-traced constants).\n\n   D-039 repair details:\n   Bound: ψ(w) = log w − 1/(2w) + r(w), |r(w)| ≤ 1/|w|² for Re w ≥ −5/4 (or −3/2), |Im w| ≥ 5.\n   \n   Piece 1 (Binet, Re W ≥ 1/2, |Im W| ≥ 5): r(W) = −2∫₀^∞ t/((t²+W²)(e^{2πt}−1))dt. Wait — need to double-check Binet's second formula for ψ: Binet's formulas for log Γ: log Γ(z) = (z−½)log z − z + ½log 2π + 2∫₀^∞ arctan(t/z)/(e^{2πt}−1)dt. For ψ = d/dz log Γ: ψ(z) = log z − 1/(2z) − 2∫₀^∞ t/((t²+z²)(e^{2πt}−1))dt. Yes (standard). So r(W) = −2∫... and |r(W)| ≤ 2∫₀^∞ t/(|t²+W²|(e^{2πt}−1))dt.\n   \n   With W = U+iV, U ≥ 1/2, |V| ≥ 5: |t²+W²|² = (t²+U²−V²)² + 4U²V²t².\n   Split at t₀ = |V|/2:\n   - t ≤ t₀: |t²+U²−V²| ≥ V² − t² − U² ≥ V²(1 − 1/4) − U² = 3V²/4 − U². Hmm U could be large? In our application U = Re w + 2 ∈ [−5/4+2, 2+2] = [3/4, 4]. So U ≤ 4, V ≥ 5: 3V²/4 − U² ≥ 0.75V² − 16 ≥ 0.75V²(1 − 16/(0.75·25)) ≥ 0.75V²·0.147?? at V=5: 0.75·25 = 18.75 − 16 = 2.75 — barely positive. Weak. Better split: t₀ = |V|/√2: t² ≤ V²/2 ⟹ |t²+U²−V²| ≥ V²/2 − U²... same U issue at V=5: 12.5−16 <0! Ugh — U up to 4 and V = 5: t² + W² at t near |V|: |W²+t²| can be as small as ~2U|V|t-ish when t² ≈ V²−U². The 4U²V²t² term saves: |t²+W²| ≥ 2U|V|t ≥ |V|t (U ≥ 1/2). At t = 5: ≥ 25. So globally: |t²+W²| ≥ max(|V²−U²−t²|, 2U|V|t) ≥ |V|t (since U ≥ 1/2). Use |t²+W²| ≥ |V|t for ALL t — wait 2U|V|t with U ≥ 1/2 gives ≥ |V|t. But we need |Im w| ≥ 5 → |V| = |Im W| = |Im w| ≥ 5.\n     Then ∫₀^∞ t/(|V|t(e^{2πt}−1))dt = (1/|V|)∫₀^∞ dt/(e^{2πt}−1) = (1/|V|)·(1/24)·(2π)²/... ∫₀^∞ dt/(e^{2πt}−1) = π²/6/(2π)² = 1/24. Hmm: ∫₀^∞ t^{s−1}/(e^{2πt}−1)dt = Γ(s)ζ(s)/(2π)^s. s=1: Γ(1)ζ(1)?? ζ(1) diverges! ∫ dt/(e^{2πt}−1) diverges at t=0 (like 1/(2πt)). So that bound diverges — must keep t in numerator: use |t²+W²| ≥ max(|V|t, c(t)) — combine: for the integral with numerator t: split at t₀ = 1:\n     - t ≤ 1: |t²+W²| ≥ |W|² − 1 ≥ 0.96|W|² (as before): ∫₀¹ t dt/(e^{2πt}−1) < 1/24 (s=2: Γ(2)ζ(2)/(2π)² = (π²/6)/(4π²) = 1/24 ✓).\n     - t ≥ 1: |t²+W²| ≥ |V|t: ∫₁^∞ t/(|V|t(e^{2πt}−1))dt = (1/|V|)∫₁^∞ dt/(e^{2πt}−1) ≤ (1/|V|)·e^{−2π}/(2π(1−e^{−2π})) ≈ (1/|V|)·3.0e-4.\n     Total: |r(W)| ≤ 2[1/(24(|W|²−1)) + 3.0e-4/|V|].\n     For |V| ≥ 5 and |W| ≥ 5: ≤ 2[1/(24·24) + 3.0e-4/5] = 2[0.001736 + 0.00006] = 0.0036. We need ≤ 1/|W|² = 1/25 = 0.04 at |W|=5 — OK! But asymptotically: the 3.0e-4/|V| term ~ 6e-4/|W| — that's O(1/|W|), TOO WEAK asymptotically. Need O(1/|W|²). Improve the tail: for t ≥ 1, split again at t₁ = |V|/2:\n     - 1 ≤ t ≤ |V|/2: |t²+W²| ≥ |(V²−U²)−t²| ≥ (V²−U²)−V²/4 = 3V²/4−U². With U ≤ 4 (application), |V| ≥ 5: ≥ 3V²/4 − 16 ≥ 3V²/4(1 − 64/(3V²)) ≥ (3/4)V²·(1−64/75) = (3/4)V²·0.147 — at V = 5 that's 2.76; for larger V better. Hmm — need a bound valid for all U ≥ 1/2 or restrict to U ≤ 4? The contour application: w = s/2 with σ ∈ [−5/2, 3] say, so U = σ/2 + 2·... wait W = w + 2, w = s/2: Re w ∈ [−5/4, 3/2+...]. Let me get concrete: the R-bound needs ψ(s/2) and ψ(s₀/2) for s with σ ∈ [−5/2, 2]. Re w = σ/2 ∈ [−5/4, 1]. After recurrence +2: Re W = σ/2 + 2 ∈ [3/4, 3]. So U ∈ [3/4, 3], |V| ≥ 5.\n     With U ≤ 3: 3V²/4 − U² ≥ 3·25/4 − 9 = 9.75 at |V| = 5, and for large V ~ 3V²/4. So |t²+W²| ≥ (3V²/4 − 9) ≥ 0.39V²·... for |V| ≥ 5: 3V²/4−9 ≥ (3/4−9/25)V² = 0.39V². Then ∫₁^{|V|/2} t/(0.39V²(e^{2πt}−1))dt ≤ (1/(0.39V²))·(1/24) = 0.107/V².\n     - t ≥ |V|/2: |t²+W²| ≥ 2U|V|t ≥ |V|t (U ≥ 3/4 ≥ 1/2·... 2U ≥ 1.5 so ≥ 1.5|V|t). ∫_{|V|/2}^∞ t/(1.5|V|t(e^{2πt}−1))dt = (1/(1.5|V|))∫_{|V/2}^∞ dt/(e^{2πt}−1) ≤ (1/(1.5|V|))·e^{−π|V|}/(2π) — exponentially small.\n     Total: |r(W)| ≤ 2[1/(24(|W|²−1)) + 0.107/V² + exp] ≤ 2[1/(24·0.96|W|²) + 0.107/V²] + ... With |V|² = |W|² − U² ≥ |W|²(1 − 9/25) = 0.64|W|²: 0.107/V² ≤ 0.167/|W|². So |r(W)| ≤ 2[0.0434 + 0.167]/|W|²·... wait: 1/(24·0.96|W|²) = 0.0434/|W|². Sum: 2(0.0434+0.167) = 0.42/|W|². Plus exp-small. So |r(W)| ≤ 0.43/|W|² for U ∈ [3/4, 3], |V| ≥ 5. (Asymptotic coefficient must be ≥ 1/12 = 0.083 — 0.43 is fine, and the bound |r| ≤ 1/|w|² target has room.)\n     \n   Piece 2 (recurrence): r(w) = r(w+2) + E(w), E(w) = log(1+2/w) − 1/(2(w+2)) − 1/(2w) − 1/(w+1). Need |E(w)| ≤ C/|w|²... E(w) ~ −1/(3w³) + O(1/w⁴) — actually E is O(1/|w|³), so any C/|w|³-ish bound works. Need explicit: |E(w)| ≤ c₃/|w|³ with small c₃ for |w| ≥ 5, |Im w| ≥ 5 (Re w ≥ −5/4).\n   f(u) = log(1+2u) − u/(2(1+2u)) − u/2 − u/(1+u), f(u) = −u³/3 + u⁴ − 2.6u⁵ + ... Need max |g(u)| = |f(u) + u³/3| on |u| ≤ 1/5 (|w| ≥ 5). Compute rigorously: the series coefficients: g(u) = Σ_{k≥4} c_k u^k with c_k = (−1)^{k+1}2^k/k − (−2)^{k−1}/2 − (−1)^{k−1} for k ≥ 2... wait need f's coefficients: log(1+2u) = Σ_{k≥1} (−1)^{k+1}2^k u^k/k. −u/(2(1+2u)) = −½Σ_{j≥1}(−2)^{j−1}u^j = Σ_{j≥1}(−1)^j 2^{j−2} u^j. −u/(1+u) = Σ_{j≥1}(−1)^j u^j. −u/2 separate. So c_k (k ≥ 2): (−1)^{k+1}2^k/k + (−1)^k 2^{k−2} + (−1)^k = (−1)^k[−2^k/k + 2^{k−2} + 1]. |c_k| ≤ 2^k/k + 2^{k−2} + 1. On |u| ≤ 1/5: |g(u)| ≤ Σ_{k≥4}|c_k||u|^k ≤ Σ(2^k/k + 2^{k−2}+1)/5^k ≤ Σ(2/5)^k(1/k + 1/4) + Σ5^{−k} = geometric: (2/5)⁴/(1−2/5)·(1/4+1/4)... ≤ (2/5)⁴·(1/2)/(3/5)·... compute: Σ_{k≥4}(2/5)^k = (2/5)⁴/(3/5) = 0.0256/0.6 = 0.0427. Times (1/4 + 1/4) [since 1/k ≤ 1/4 for k≥4, plus the 2^{k−2} contributes (1/4)2^k] = 0.0427·0.5 = 0.0213; plus Σ_{k≥4}5^{−k} = 5^{−4}/(4/5) = 0.002. Total ≤ 0.0233. So |g(u)| ≤ 0.0233 for |u| ≤ 1/5. Since g(u) = O(u⁴), Schwarz lemma: |g(u)| ≤ 0.0233·(5|u|)⁴ = 14.6|u|⁴. Hmm — at |u| = 0.2: 14.6·0.0016 = 0.0233 (consistent, no gain at boundary). Better: bound directly |E(w)| = |f(u)| ≤ |u|³/3 + 0.0233... at |u| = 0.2: 0.00267 + 0.0233 = 0.026. Or with the |u|⁴ factor on the whole region |u| ≤ 0.2 — the max of |g| might be interior... the direct sum bound 0.0233 is for |u| ≤ 0.2 — to get |g(u)| ≤ M|u|⁴ use max-modulus on g(u)/u⁴: max over |u|≤0.2 of |g/u⁴| ≤ max on boundary |u|=0.2: |g|/0.0016 ≤ 0.0233/0.0016 = 14.6. So |g(u)| ≤ 14.6|u|⁴, giving |E(w)| ≤ 1/(3|w|³) + 14.6/|w|⁴. At |w| = 5: 0.00267 + 14.6/625 = 0.00267+0.0234 = 0.026. Times |w|² = 25: 0.65. Hmm — so |E(w)| ≤ 0.65/|w|² at |w| = 5, decaying as 1/(3|w|³) + 14.6/|w|⁴ ~ (1/3)/|w|³ for large w. \n   Assembly: |r(w)| ≤ |r(W)| + |E(w)| ≤ 0.43·(1/|W|²) + 0.65/|w|²·(at worst). |W| = |w+2| ≥ |w|−2 ≥ (3/5)|w| for |w| ≥ 5 ⟹ 1/|W|² ≤ 25/(9|w|²) = 2.78/|w|². So |r(w)| ≤ (0.43·2.78 + 0.65)/|w|² = (1.19 + 0.65)/|w|² = 1.84/|w|². EXCEEDS the target 1/|w|²! At |w| = 5. Hmm. Need sharper: the worst case is |w| = 5 with Re w = −5/4 (then W = w+2 has U = 3/4, |W| = |0.75+5i| = 5.056, 1/|W|² = 0.0391; the (3/5) slack is nearly tight). Let's recompute more carefully at the actual worst point instead of the lossy 25/9 factor: |r(w)| ≤ 0.43/|W|² + 1/(3|w|³) + 14.6/|w|⁴. At |w|=5, |W| ≥ 5.056: 0.43/25.56 = 0.0168 + 0.00267 + 0.0234 = 0.0429 vs 1/|w|² = 0.04. Just over! Marginal. Improvements available: (a) the U ≤ 3 restriction gives better |W|: |W|² = (U)²+V² with U ≥ 3/4: |W|² ≥ 0.5625 + V², and |w|² = (U−2)² + V² ≤ ... at V = 5, U = 3/4: |w|² = 1.5625+25 = 26.56 vs |W|² = 25.56 — ratio 1.039, not 25/9. Generally |w|²/|W|² = ((U−2)²+V²)/((U)²+V²) ≤ for U ∈ [3/4, 3], V ≥ 5: max at U = 3/4, V = 5: 26.56/25.56 = 1.039; at U=3,V=5: (1+25)/(9+25) = 0.765. So ≤ 1.04. (b) tighten 0.43 — at the specific point U = 3/4, V = 5: the 3V²/4 − U² = 18.75−0.5625 = 18.19 term: 1/(24(|W|²−1)): |W|² = 25.56: 1/(24·24.56) = 0.0017; 0.107/V² → with the actual numbers: 1/(0.39·25)·(1/24)... wait I already folded. Let me just redo the tail bound with U ≥ 3/4 (not 1/2) and U ≤ 3:\n     - middle piece 1 ≤ t ≤ |V|/2: |t²+W²| ≥ 3V²/4 − 9 (U ≤ 3). At V = 5: 9.75. Generally ≥ V²(3/4 − 9/V²) ≥ V²·0.39 (V≥5). ∫ ≤ 1/(0.39V²)·(1/24) = 0.107/V².\n     - 0 ≤ t ≤ 1: |t²+W²| ≥ |W|²−1 = U²+V²−1 ≥ V²−... ≥ 0.96|W|² (|W| ≥ 5): ≤ 1/(24·0.96|W|²) = 0.0434/|W|².\n     - tail t ≥ |V|/2: ≤ exp.\n     |r(W)| ≤ 2(0.0434/|W|² + 0.107/V²) ≤ 2(0.0434 + 0.107/0.64)/|W|²·... hmm 0.107/V² with V² ≥ 0.64|W|² (U ≤ 3, |W| ≥ 5: U² ≤ 9 ≤ 0.36|W|²): ≤ 0.167/|W|². Total 2(0.0434+0.167) = 0.421/|W|².\n     At the worst point (|w| = 5, U = 3/4, V = 5): 0.421/25.56 = 0.01647. E: 0.026. Sum: 0.0425 vs 0.04 allowed. STILL marginally over. \n   Options: (i) relax the claim to |r(w)| ≤ 1.1/|w|² — does the R-bound care? The digamma bound enters the ψ-difference in the reference-point assembly as O(1/|t|²) terms — utterly negligible either way (constants absorbed in B₀). So simply claim |r(w)| ≤ 1.1/|w|² or even 2/|w|² — the application is insensitive! The original \"1/|w|²\" was aesthetic. Actually simpler: claim |r(w)| ≤ 2/|w|² for Re w ≥ −5/4, |Im w| ≥ 5 — assembled 1.84 ≤ 2 with margin, rigor via the lossy path. OR (ii) shrink the E-majorant: the 14.6|u|⁴ is lossy; direct: |E(w)| ≤ |u|³/3 + 0.0233 — at |w|=5: 0.026. The dominant loss is the 0.0233 absolute (vs true max ~0.01?). Meh — simplest rigorous path: use the lossy assembly → |r(w)| ≤ 1.84/|w|², claim 2/|w|². The downstream use: |ψ(s₀/2) − ψ(s/2)| ≤ |log(s₀/s)| + |1/(2s₀) − 1/(2s)| + 2/|w₀|² + 2/|w|² = O(1/|t|) — any constant works.\n\n   Hmm wait, actually — where exactly is the digamma bound needed in the NEW reference-point derivation? Only in the ψ-difference, where any O(1/|w|²) constant suffices. So the repair is easy. But note: the ORIGINAL D-036 (non-reference-point) assembly used ψ(s/2) = log(s/2) + O(1) — the log term there combined with the far-sum... whatever, superseded.\n\n   Also should double check Binet formula applicability: Binet's second formula for ψ holds for Re z > 0. W = w+2 has Re W ≥ 3/4 > 0 ✓.\n\n2. **4.5b** (symbolic concavity): B(n) components as functions of n ≥ 20 with T = √n + 1. Components: 2N(T), 4T[(2/3)C₀(T) + A log(T+2) + B₀], 40M₀ (const), 2(e⁴+1)[7H(T) + 3.3], I_I tiny, +1. C₀(t) = 0.433 log(t+2) + 2.5. H(T) = 3C₀(T)² + A log(T+2) + B₀.\n   Write each as function of T (T = √n+1, n ≥ 20 ⟹ T ≥ 5.47): margin(n) = S_∞^lower(n) − B(n) + 1. Convexity in n: margin″ > 0. Since T is concave in n (T″ = −1/(4n^{3/2})), composition f(T(n)): d²/dn² = f″(T)·(T′)² + f′(T)·T″. T′ = 1/(2√n) > 0, T″ < 0. For f increasing convex in T: f″(T′)² ≥ 0 but f′T″ < 0 — sign unclear. The pieces: 2N(T) ~ (T/π)log(T/2πe) — convex increasing in T; in n: (T/π)log T with T = √n: ~ √n log n /(2π) — concave in n (verified earlier). Symbolic check: compute second derivatives of each component with sympy, show margin″(n) > 0 for n ≥ 20 (bound each piece's second derivative: B_i″(n) ≤ 0 or small, (S_∞)″ = 1/(2n) > dominant). Doable: write a sympy script producing exact symbolic second derivatives and prove negativity/positivity via bounds. This is E-028 or fold into a script. Plan: derivation D-057 with a sympy-assisted symbolic check (script + output). \n\n3. **4.5c** (valley lower bounds) — the hard one. Repair route: node-killing test functions. Idea: in (I1)'s sup, use v with v̂ = (2a)sinc-profile × Π(1 − ((z−γ)/(γ_j−γ))²)-ish nodes at the J nearest zeros. Concretely: choose v̂(z) = 2a·sinc((z−γ)a)·Π_{j=1..J} ((z−γ_j)/(γ−γ_j))²?? But v̂ must be PW of type a (band-limited) — multiplying sinc by polynomials breaks the type. Alternative: work in x-space: v = v₀·(1 + Σ c_j e^{iω_j x})-type corrections, or project: the sup over PW-type-a functions of |v̂(γ)|²/(‖v‖² + Σ|v̂(γ')|²) — the optimizer is the reproducing kernel combination: v̂ = Σ_j c_j K_a(·, z_j)... For the LOWER bound we can use v̂(z) = K_a^{finite}(z, γ) − Σ_{j≤J} α_j K_a^{finite}(z, γ_j) with K_a^{finite}(z,w) = sin(a(z−w))/(π(z−w)) (PW reproducing kernel) and α chosen to zero out v̂ at the J nearest other zeros. Then numerator |v̂(γ)|² = (a/π − Σ α_j K(γ,γ_j))²... this is exactly finite-dimensional linear algebra with explicit error from the remaining zeros (Lemma D handles the tail since v̂ decays 1/(z−·)). The nearest-neighbor terms are killed exactly ⟹ the C₀a²/4 loss disappears ⟹ floor 1 − O((1+log γ)/a²)-ish plausible. This is genuine math work — feasible but the longest item. The B3 fix (sinc²((z−γ)a)) is trivial; m_γ fix trivial; δ_crit restatement (bare-form indefiniteness for δ > 0) — provable via strict Schwarz + background smallness in the difference direction... hmm, \"background cannot cover\" needs the background value in the difference direction < pair's negative eigenvalue — that's a real estimate: background(diff) ~ O(δ²·C₀-ish), pair negative eigen ~ −O(δ²·K)... both O(δ²) — constants matter; provable with kernel asymptotics (K_a(γ*±iδ) explicit via Christoffel–Darboux/Dirichlet kernel). Medium difficulty.\n\n4. **4.5d** — write-ups (L1, form domain, punctured compacts, R_τ distributional). Paper-work, no new math except form-domain identification (zero-side domain = operator-form domain — needs the explicit formula on the right class; standard closure argument).\n\nThat's a big stage. Pacing directive: sub-stages, each ending clean/committed. Also the 5 PM break prompt rule. Current time ~15:50? (The session's been running ~40+ min based on timestamps 15:10 log.) Plan sub-stages:\n\n- Sub-stage 1 (now): **4.5a** — D-056 (reference-point rewrite + D-039 repair as D-039 replacement or D-056 includes both) + E-028 (margin rerun with traced constants → new N₀) + R-004 tag restoration. Commit.\n- Sub-stage 2: **4.5b** — D-057 symbolic concavity + E-verification. Commit.\n- Sub-stage 3: **4.5c** — valley lower-bound repair (D-058) — the research-level item. Commit.\n- Sub-stage 4: **4.5d** — bookkeeping write-ups (D-059). Commit.\n\nStart 4.5a. Write D-056 with the full derivation, carefully. Then E-028: modify E-023 script constants (A, B₀, I_I length, certified ζ′/ζ(2)) and rerun the margin computation to get new N₀ — need to check E-023 script structure first.\n\nLet me look at E-023 script to understand what to parameterize.","signature":""},{"type":"text","text":"Proceeding with the repair stage. Order per STATE.md: **4.5a → 4.5b → 4.5c → 4.5d**, each as a clean committed sub-stage."},{"type":"tool_use","id":"TodoWrite_139","name":"TodoWrite","input":{"todos":[{"activeForm":"Repairing R-bound + digamma (4.5a)","content":"4.5a: D-056 reference-point R-bound rewrite + D-039 digamma repair + I_I housekeeping; E-028 margin rerun with traced constants; restore R-004 tag","status":"in_progress"},{"activeForm":"Proving margin concavity (4.5b)","content":"4.5b: symbolic concavity proof of B(n) components (D-057 + script)","status":"pending"},{"activeForm":"Repairing valley lower bounds (4.5c)","content":"4.5c: valley-theorem lower bounds repair + B3/m_γ fixes + detector restatement (D-058)","status":"pending"},{"activeForm":"Writing bookkeeping items (4.5d)","content":"4.5d: L1 write-up + form-domain identification + R-003 bookkeeping (D-059)","status":"pending"},{"activeForm":"Updating paper + session end","content":"Paper (draft.tex) tag corrections + session end protocol","status":"pending"}]},"extra_content":null}]