# Fixed-endpoint discrepancy: the split at x^(1/2-eps'), the below-level Type I piece, and the remainder that is left **Twin-prime infinitude, D^(e_1)>=-4x/25+o(x) and every sufficient signed margin remain OPEN. This note estimates one piece (the below-level Type I piece, corrected and independently reviewed on 2026-09-09) and exhibits the rest with exact sums, trivial sizes and the first unmatched hypothesis per source.** It is lane A's second checkpoint of 2026-09-08, dispatched from e90d49b after the truncation T^top=O_(A,eps)(x/log^A x) of [centered-discrepancy-estimate.md](centered-discrepancy-estimate.md) section 3a was accepted ([research-round-validation.md](research-round-validation.md) section 7). Lane / stable question id: A (second checkpoint) / Q-fixed-endpoint-discrepancy Starting commit / report commit or shared-checkout paths: e90d49b, shared checkout; research/fixed-endpoint-discrepancy.md, research/fixed-endpoint-discrepancy-validation.js Disposition / exact claim / unproved hypotheses: exact three-piece reduction from accepted inputs; T_I^low=O_(A,eps')(x/log^A x), corrected and accepted after the independent 2026-09-09 reading (g truncated at (log x)^L, both tails paid); B (Type II below level plus the band) unestimated: the recorded parity object Changed step compared with the reviewed baseline: the fixed-endpoint object is split at e_0=floor(x^(1/2-eps')) and the cofactor Mobius is decomposed by Vaughan's identity; the density projection of both parts is evaluated; the consumer is restated as S=C_2x+B+o(x); after V4, the coprimality expansion in the Type I piece is truncated at g<=(log x)^L so that every BV modulus is at most x^(1/2-eps'/3)(log x)^L, and the two g-tails are bounded in section 4.1 Source theorem and first unmatched hypothesis, if any: none imported beyond (BV*), (3a.9), PNT; for the band, every source in section 3 fails at absolute values over all moduli near x^(1/2+eps') in one fixed class, or at the signed weight Validation command, falsifier, result and compute used: node research/fixed-endpoint-discrepancy-validation.js (0.5 s, one core); exact identities at x=2^10..2^16 pass, deletion controls fire, density and multiplicity formulas checked finitely; no asymptotic step is tested Independent reviewer / disposition: the 2026-09-09 integration review reconstructs the repaired tails, density, multiplicity and uniform-mean application; accepts (4.1) with the bookkeeping corrections below. The original q<=e_0UV claim remains refuted by the retained witness. See research-round-validation.md section 12 (consumer in 12a). Full-consumer payoff and unpaid complement: none; the unpaid complement is B, of elementary size O(x log^5 x), required >= -(C_2-1/200)x+o(x) Proposed shared-record changes / next bounded obligation: section 7; section 8 ## 1. Question, disposition and what remains open Question: split D^(e_1) at a level e_0=x^(1/2-eps'), say what estimate the below-level part needs for f(n)=Lambda(n-2)mu(n) in the progressions n=0 (mod e) on the fixed interval, name the piece an existing theorem estimates, define the band count with its density and clipped endpoints, and give one bounded attempt. Disposition. The split, the density projections and the Vaughan three-piece decomposition are exact and finitely checked. The below-level Type I piece is estimated in section 4.1 by the accepted inputs of the truncation proof; second reader V4 found its first version's modulus bound false, the correction (truncation of the coprimality divisor with both tails paid) was applied on 2026-09-08, and the corrected derivation was independently reviewed and accepted on 2026-09-09 with the bookkeeping corrections below. Everything else is exhibited: the below-level Type II piece and the band together form one sum B with S(x)=C_2x+B(x)+O_A(x/log^A x). B is a shifted-prime bilinear remainder carrying one Mobius factor on a long variable, the same family as B(x) of [prime-detection-spec.md](prime-detection-spec.md) (13) and the residual of [shifted-prime-decomposition.md](shifted-prime-decomposition.md), with different coefficients and ranges. No theorem with matched hypotheses estimates any part of it; the stop rule of the assignment applies and the pass stops at the exact sum. Open: D^(e_1)>=-4x/25+o(x), equivalently B+2C_2M>=-4x/25+o(x). Using M<=A_2 x/2 and C_2(1-A_2)>33/200, this implies B>=-(C_2-1/200)x+o(x). The converse does not follow: a lower value of M can violate the D margin while that B bound holds. Both are sufficient for infinitude on unbounded dyadic scales; the general fixed-c_0 consumer (H_B) is also OPEN. The regression is a logical counterexample to the claimed equivalence, not an arithmetic counterexample to either target. ## 2. Exact statement Notation of the handoff and of centered-discrepancy-estimate section 1: x=2^j, J=(x/2,x], Lambda standard (prime powers included), f(n)=Lambda(n-2)mu(n), M=sum_{n in J} f(n), F_1:=sum_{n in J} f(n) log n, C_2 the twin constant, fixed 0=(1/2-eps'-o(1))log x, |M|=O(x) and |F_1|=O(x log x), \[ Q_{\rm low}=-2C_2M+O_A(x\log^{1-A}x),\qquad Q_{\rm band}=O_A(x\log^{1-A}x). \tag{2.2} \] Hence, exactly up to these accepted errors, \[ \mathcal D^{(e_1)}=2C_2M+P_{\rm low}+P_{\rm band}+O_A(x\log^{1-A}x). \tag{2.3} \] ### 2.3 The band count, its density and the clipped endpoints Two exact arrangements of P_band, both used in section 4.3. *Modulus e, cofactor m.* For odd squarefree e in [e_0,e_1) define \[ N(e):=\sum_{\substack{n\in J\\ n\equiv0\ (e)}}f(n)\log\frac ne =\mu(e)\sum_{\substack{m\in I_e\\ (m,e)=1}}\mu(m)\log m\,\Lambda(em-2),\qquad \kappa(e):=\frac{F_1-M\log e}{\varphi(e)}, \] so that D_band=-sum_{e_0<=e=x+1 and e_0m<=x/2+1. Otherwise use the displayed max/min endpoints, with any empty interval contributing zero. Expanding mu^2(n/m)=sum_{b^2|n/m}mu(b) with b odd and 1_((n/m,m)=1)=sum_{g|(n/m,m)}mu(g) as in section 3a.1, the inner sum is sum_{b,g}mu(b)mu(g) sum_{n in I_m, n=0 (m[b^2,g])}Lambda_0(n-2) plus the single power-of-two atom, with moduli q=m[b^2,g] up to 2x^(1/2+eps')(log x)^(3L) after the truncation b,g<=(log x)^L of 3a.3. Its main term is |I_m| sum_{b,g}mu(b)mu(g)/phi(m[b^2,g]), the density of 3a.2; this is the "kappa" of the modulus arrangement, and |I_m| is monotone on each of the three m-ranges. Trivial size of the band. Counting multiples in (2.1) and using Lambda<=log x gives O_(eps,eps')(x log^3 x). Brun–Titchmarsh improves this to O_(eps,eps')(x log^2 x) for fixed eps<1/50, since e=1 ([mobius-bv-derivation.md](mobius-bv-derivation.md), checked exactly to 10^5 there and to 4096 here): \[ \mu=\mu_{>U}*\mu_{>V}*1-\mu_{\le U}*\mu_{\le V}*1+\mu_{\le U}+\mu_{\le V}. \] For ex/(2e_0)>=x^(1/2+eps')/2>max(U,V), so the two linear terms vanish. Put c(r):=sum_{dk=r, d<=U, k<=V}mu(d)mu(k), |c(r)|<=tau(r), supported on r<=UV, and gamma_V(b):=sum_{k|b, k>V}mu(k), |gamma_V(b)|<=tau(b). Then exactly \[ P_{\rm low}=T_I^{\rm low}+T_{II}^{\rm low}, \tag{2.5} \] \[ T_I^{\rm low}=\sum_{\substack{eU,\ b>V,\ ab\in I_e\\ (ab,e)=1}}\mu(a)\gamma_V(b)\log(ab)\,\Lambda(eab-2). \tag{2.7} \] Ranges in (2.7): a in (U, x/(eV)], b in (V, x/(eU)]; so a runs from x^(eps'/3) up to x^(1-eps'/3) as e and b vary. Heath-Brown's identity for mu (the 1/zeta analogue of his identity for Lambda, K pieces mu_{<=Z}^{*j}*1^{*(j-1)}) yields the same two classes with a movable boundary; the balanced region, both factors of the cofactor near a fixed power, is present under every such identity, as [heath-brown-edges.md](heath-brown-edges.md) records for Lambda. It is not used further here. **What the below-level part needs.** In the flipped form, the twisted sequence f(n) in the progression n=0 (mod e) on the fixed interval J is the cofactor sum sum_{m in I_e,(m,e)=1}mu(m)log m Lambda(em-2). After (V), the Type I part asks for primes n-2 in the reduced class -2 modulo e[r,g]<=e_0UVG=x^(1/2-eps'/3)(log x)^L, after truncating the coprimality divisor at g<=G (section 4.1), with the signs carried by c(r) and evaluated through Mobius means (section 4.1); the Type II part asks for the bilinear sum (2.7). No theorem for f itself in progressions is used. ### 2.5 The reduction and the consumer Combining (12), (3a.1), (2.3) and (2.5), \[ S(x)=C_2x+T_I^{\rm low}(x)+\mathcal B(x)+O_A(x\log^{1-A}x),\qquad \mathcal B:=T_{II}^{\rm low}+P_{\rm band}, \tag{2.8} \] with the accepted inputs only. Written on n=eab with b=1 in the band, \[ \mathcal B(x)=-\sum_{n\in J}\Lambda(n-2)\,W(n),\qquad W(n)=\sum_{\substack{eab=n,\ e\ \mathrm{odd},\ \mu^2(e)=1\\ (ab,e)=1}}\mu(a)\log(ab) \Bigl[\mathbf 1_{eU}\gamma_V(b)+\mathbf 1_{e_0\le e=-4x/25+o(x) is B+2C_2M>=-4x/25+o(x); - the direct twin consumer S>=x/200+o(x) is B>=-(C_2-1/200)x+o(x); - the first implies the second through 2C_2M<=C_2A_2x+o(x), which is how the fixed fraction 4/25 was obtained in moving-cutoff-parity (14)-(16). **Named hypothesis (H_B), all quantifiers.** With eps, eps', U, V, e_0, e_1 as fixed above and B as in (2.9): there exist c_0>0 and an unbounded set of j such that B(2^j)>=-(C_2-c_0)2^j+o(2^j). This is a conditional sufficient statement, not a theorem; under section 4.1 it is equivalent to S(x)>=c_0x+o(x) on those scales. The weaker margins of handoff section 3 (c_0x/log^K x) transfer verbatim. For a safe elementary size comparison, (2.9) gives |T_II^low|<=log x sum_(n in J) Lambda(n-2)tau_4(n) <U; here mu(n) is decomposed through its cofactor and Lambda(n-2) is untouched. The two remainders B(x) and B differ in coefficients and ranges, not in kind. Local records read before this pass: OUTCOMES and QUESTIONS entries for Q-centered-discrepancy-estimate, Q-moving-cutoff-parity, Q-consumer-comparison, Q-mobius-bv-derivation, Q-shifted-prime-decomposition, Q-prime-detection-inputs, Q-heath-brown-edges, Q-shifted-prime-mobius-sums; SEARCH-CONVENTIONS rows for the band of D_y and for the twisted sequence. No new literature search was run; the source matrix of centered-discrepancy-estimate section 4 was matched against the actual band objects of section 2.3 and 4.3 here. | Source, locator (as read in the existing matrix) | Statement it makes | Band object it is matched against | First unmatched hypothesis | |---|---|---|---| | Tao, Notes 3, Theorem 17, prefix form (BV*) | absolute values over all q<=x^(1/2)log^-B x, reduced classes, ineffective | modulus arrangement (2.4), moduli m[b^2,g] | level: the band moduli reach 2x^(1/2+eps')(log x)^(3L); below x^(1/2-eps'/3) it is used in section 4.1 | | [Maynard I, arXiv:2006.06572v2](https://arxiv.org/abs/2006.06572) Theorem 1.1 | absolute values over q_1q_2 with Q_1Q_2^2x fails; with Q_1 the band factor, Q_1^12>x^6>x^4 fails; Cor. 1.2 admits a divisor in (x^(2eps'+eta), min(x^(1/10-7eps'/5-eta),x^(1/2-19eps'-eta))); both its lower and upper range constraints are required. A lower bound on the smaller factor alone does not establish coverage; outside that sufficient range this corollary supplies no bound | | Maynard I, Corollary 1.3 | all but 18·delta·Q·phi(a)/a moduli in [Q,2Q], Q=x^(1/2+delta), absolute values | modulus arrangement, all m in a dyadic block | the exceptional moduli carry, with the log weight, trivial mass of order 18·delta·x per block; summed over the blocks delta in (0,eps'] this is of order eps'^2 x log x, above O(x); the signed weight mu(m) on the exceptional set is the obstruction, as recorded | | BFI II Theorems 3, 5* (restated in Maynard I Lemmas 8.4-8.5; primaries unread) | absolute values over q~Q in (x^(1/2)log^-A x, x^(2/3-e)) for triple convolutions of the prime variable with range constraints | (2.9) read as the sequence n=p+2 with a triple-convolution weight e·a·b | the sequence here is Lambda(n-2) itself in the progression, not a convolution; the convolution sits on the modulus side, and the bad shapes of Maynard I section 3.2 are uncovered in any case | | BFI II + III Theorem A (as in Maynard I section 1.1; primaries unreached) | no absolute values, Q=x^(1/2+delta), fixed a: sum_{q~Q}(pi(x;q,a)-pi(x)/phi(q))=O(delta^2 x/log x+x(log log x)^O(1)/log^3 x) | modulus arrangement after Vaughan on mu(m): Type I pieces with modulus r·s·b^2g, s unweighted in a block, fixed class -2 | shape matches only for s unweighted and for all q in [Q,2Q], not for multiples of r·b^2g; and a delta^2 saving per block leaves, after the log weight and the (eps+eps')log x/log 2 blocks, order eps'^3 x log x·(log UV)^2 from sum_r tau(r)/r, above O(x); the signed c(r), mu(b), mu(g) are then summed in absolute value | | BFI I Theorem 10, Maynard II Theorem 1.1 | well-factorable (triply well-factorable) lambda_q, fixed a, level x^(4/7-e) (x^(3/5-e)) | modulus arrangement: lambda_q=mu(m)log m·1_{m in range} or its Vaughan pieces 1_{r|m}log m | not well-factorable (recorded); the Type I piece 1_{r|m}·1_{m~Q} is a convolution of an indicator with an indicator of a long range, which is not a factorization into 1-bounded pieces of every prescribed pair of supports | | [Polymath, arXiv:1402.0811v3](https://arxiv.org/abs/1402.0811) Theorem 1.1 | x^delta-smooth squarefree moduli, level 1/2+7/300 | modulus arrangement | the band moduli m are arbitrary squarefree; the smooth sub-family carries no sign advantage | | [Drappeau, arXiv:1504.05549v4](https://arxiv.org/abs/1504.05549), Titchmarsh sum | unweighted modulus average near x^(1/2) with log-power error | Type I pieces on the modulus | window of log-power width around x^(1/2) only; the band has power width | | Murty–Vatwani Theorem 1.1, EH_{mu_2}(x^(1/2+eps)) | hypothesis, all classes, all prefixes | D^(e_1) directly | it is a hypothesis; (H_B) is one-sided and one-class, weaker, and unproved | UNREAD in this pass: BFI I, BFI III, Fouvry 1985 primaries (unreached on 2026-09-08 per the existing matrix); no new fetch was attempted, since no row's shape matched before its first hypothesis. ## 4. Proof of the Type I estimate and the decisive exhibited remainder ### 4.1 Proposition (corrected and reviewed 2026-09-09) V4 found that the first version bounded the modulus q=e[r,g] by e_0UV, which is false for g>1: g runs over every divisor of e and [r,g]>=g, so q reaches e^2 at g=e, r=1 (at x=2^16, e_0=212, U=V=3: e=g=209, r=1 gives q=43681>1908; validator section 7). The handler's first reading did not catch this. The repair below truncates g at G:=(log x)^L, exactly as 3a.3 truncates b and g, and pays both tails. The 2026-09-09 independent reading accepts the repair after correcting the log powers and atom bound below. For every fixed A>0 and 0=x_0(A,eps'), \[ T_I^{\rm low}(x)=O_{A,\epsilon'}\bigl(x\log^{-A}x\bigr). \tag{4.1} \] Constants are ineffective (Siegel through (BV*) and (3a.9)). *Step 1: powers of two and coprimality.* Let Lambda_0 be Lambda with Lambda_0(2^a)=0. Terms of (2.6) with em-2 a power of two have em=x/2+2=:n_0 (one value, as in 3a.1). Their total is at most log 2·log x·sum_{e|n_0}sum_{r|n_0/e}|c(r)|tau(e)<=2 log 2·log x·tau(n_0)^3(UV)^(1/2)<<_delta x^(3delta+eps'/3)log x, using |c(r)|<=tau(r)<=2 r^(1/2) and tau(n)<<_delta n^delta; with delta=1/100 this is O(x^(1/4)). In the remaining terms em is odd, so e, m, r are odd. Expand 1_((m,e)=1)=sum_{g|(m,e)}mu(g); with r|m and g|m the condition is [r,g]|m, and [r,g]=rg/(r,g) (validator section 7). Fix L>0, put G:=(log x)^L and split the g-sum into the body g<=G and the tail g>G: \[ T_I^{\rm low}=T_I^{\rm body}+T_I^{\rm tail}+E_2,\qquad T_I^{\rm body}=\sum_{\substack{eG. In the body the modulus q:=e[r,g] is odd (so the class -2 is reduced) and satisfies, since [r,g]<=rg, \[ q\le e_0\,UV\,G\le x^{1/2-\epsilon'/3}(\log x)^L=:Q_0'. \tag{4.2'} \] *Step 1a: the tail on the progression side.* Each term of T_I^tail is at most log x·log x·(number of m in I_e with [r,g]|m)<=log^2x·(x/(2e[r,g])+1), and x/(e[r,g])=x(r,g)/(erg). Writing (r,g)=sum_{d|(r,g)}phi(d), \[ \sum_{r\le UV}\frac{\tau(r)(r,g)}{r}=\sum_{d\mid g}\varphi(d)\sum_{\substack{r\le UV\\ d\mid r}}\frac{\tau(r)}{r} \le\sum_{d\mid g}\tau(d)\sum_{r'\le UV}\frac{\tau(r')}{r'}\ll\tau(g)^2\log^2x, \] using tau(dr')<=tau(d)tau(r') and sum_{d|g}tau(d)=tau_3(g)<=tau(g)^2. Then sum_{g|e, g>G}tau(g)^2/g<=tau(e)^3/G (tau(g)<=tau(e), tau(e) divisors) and sum_{eG}}\sum_{r\le UV}\tau(r)\Bigl(\frac{x(r,g)}{2erg}+1\Bigr) \ll\frac{x\log^{12}x}{G}+x^{1/2-\epsilon'/3}\log^4x, \tag{4.2''} \] For the weighted count, sum_{r<=UV}|c(r)|<=sum_{r<=UV}tau(r) <G}mu(g)/phi(e[r,g]). Since phi([r,g])=phi(rg)/(r,g)>=phi(r)phi(g)/(r,g) and phi(e[r,g])>=phi(e)phi([r,g]), the tail density contributes at most, with sum_{n in J}log(n/e)<=(x/2)log x, \[ E_g^{\rm main}\le\frac x2\log x\sum_{eG}}\frac1{\varphi(g)}\sum_{r\le UV}\frac{\tau(r)(r,g)}{\varphi(r)} \ll x\log^3x\,(\log\log x)^2\,\frac1G\sum_{e=x_0(A_1,eps',L) because x^(-eps'/3) dominates every logarithmic power), \[ E_{BV}:=3\log x\sum_{q\le Q_0'}c'(q)D(q)\le3\log x\Bigl(\sum_{q\le Q_0'}\tau(q)^8D(q)\Bigr)^{1/2}\Bigl(\sum_{q\le Q_0'}D(q)\Bigr)^{1/2} \ll x\log^{1+(257-A_1)/2}x. \tag{4.7} \] *Step 6: the remaining terms of (4.3).* Using phi(e[r,g])>=phi(e)phi(r), sum_{eU}*gamma_V on m) produces the shapes in the Maynard and BFI rows of the matrix with the unmatched hypotheses recorded there; the accumulated log weight and the (eps+eps')log x/log 2 dyadic blocks defeat a constant-factor saving per block, by the count in section 2.3. In the cofactor arrangement, the band is the balanced part of B: both e and m within a factor x^(eps+eps') of the square root, one Mobius sign on m, a squarefree filter on e. It is the same family as 4.2. ### 4.4 Why the pass stops Every piece of D^(e_1) other than T_I^low and the density projections is an instance of one family: Lambda(n-2) at shift 2 against a product of a Mobius factor on a long variable and a divisor-bounded cofactor. That family is the recorded unfilled input (prime-detection-spec (14), shifted-prime-decomposition, consumer-comparison sections 1 and 5). A shift average would not select shift 2; a Type I estimate is not the consumer; the stronger absolute-value statement (4.9) would pay the band piece P_band, but no inspected source states it, and it does not touch the below-level Type II piece T_II^low of (2.7): with (4.9) the D-margin still requires 2C_2M+T_II^low>=-4x/25+o(x). A bound for the actual signed weights of B could suffice without (4.9); (4.9) cannot replace a signed bound for T_II^low. The exact sum is (2.9) and the required rate is (H_B). The pass stops here, as the assignment's stop rule directs. ## 5. Validation, falsifier and outcome [fixed-endpoint-discrepancy-validation.js](fixed-endpoint-discrepancy-validation.js) (0.5 s, one core, output embedded) checks at x=2^10, 2^12, 2^14, 2^16 with eps=eps'=1/60: the split D^(e_1)=D_low+D_band and D_R=P_R-Q_R with P_R in the flipped form (2.1) against the definition; Q_band against the two odd Mobius means; the dual divisor identity on odd n, P^(e_1)_odd=sum Lambda(n-2)Lambda(n)mu^2(n)+sum Lambda(n-2)mu^2(n)sum_{e'|n, e'<=n/e_1}mu(e')log e', with the even atom bounded; Vaughan's identity (V) at every m<=4096 for (U,V)=(3,3),(2,5),(4,6) and the three-piece split (2.5) at each pair; two deletion controls (dropping the coprimality filter, dropping the odd restriction) that change P_low at three or more scales; the density (4.4) on 4128 pairs; the multiplicity bound c'(q)<=tau(q)^4 for q<=3000; after V4, the identity [r,g](r,g)=rg for r,g<=200 and the modulus witness at x=2^16: e=g=209, r=1 gives q=43681 above e_0UV=1908, triples above e_0UV are counted, and with g<=G=8 no body modulus exceeds e_0UVG. All pass. These are finite checks of exact algebra; they test no asymptotic step, and the derivation of section 4.1 carries (4.1) alone. Falsifier for (4.1): a defect in (4.4), in the tail bounds (4.2'') and (4.4'), in c'(q)<=tau(q)^4, or in the application of (3a.9) at (k,e). The first version's modulus bound was such a defect, found by V4 and witnessed finitely. Falsifier for (2.8)-(2.9): a finite x at which the validator's split, dual or three-piece check fails; the four scales 2^10..2^16 all pass. No finite computation bears on (H_B), and no census was run (zero enumeration, per the contract). The finite ratios printed at x=2^16 are not readings of any asymptotic quantity. ## 6. Payoff for the full twin consumer None. The identity (2.8) restates the consumer in the classical Type I/II shape; the Type I piece is paid at the reviewed scope and the remainder B is the parity object. The regional statement obtained is: the below-level Type I piece of the cofactor Mobius is O_(A,eps')(x/log^A x). The conditional statement is: (H_B) implies S(x)>=c_0x+o(x) on unbounded dyadic scales, hence twin-prime infinitude, using the reviewed (4.1). Nothing here narrows the unpaid complement: B has elementary size O(x log^5 x) against a required lower bound of size -(C_2-c_0)x. ## 7. Validation and integration The lane validator remains the finite identity and original modulus-defect check. The independent 2026-09-09 controls in [research-round-validation.js](research-round-validation.js) add exact rational checks of (4.4) including nonsquarefree r and of the consumer implication. The proof above, not the finite values of T_I^low, establishes (4.1). [research-round-validation.md section 12](research-round-validation.md) owns this review; OUTCOMES and the handoff carry its current disposition. ## 8. Next move or reopening condition The decompositions inspected here leave the exact B in (2.9) unestimated. A useful continuation must supply a bound with its actual coefficients, shift 2, growing ranges and a rate that pays (H_B), or a changed representation with a quantified saving and its complement paid. Examples of stronger sufficient inputs for the band piece are (4.9) or a treatment of the exceptional moduli in the cited Maynard range that admits the Mobius weight; each leaves T_II^low, so neither closes the D-margin alone. These are not necessary conditions and do not exclude a new decomposition. A correctness concern in an accepted step is also a reason to reopen it. Revision history: research/history/CHANGELOG.md. Re-applied 2026-09-19 (job #2680, review #154): the 2026-09-16 mirror cut (history v3) re-imported the pre-#151 file; this version restores accepted return #151 (v2, 21dce4f3) unchanged except for this line.