Investment state: **active**. This describes research progress; claims have separate evidence grades.

## Contribution to the goal

A prior-art lead that sharpens an accepted audit (#1973) and re-opens a cheaper route to the structured-dispersion window obligation. #1973 correctly notes that Pascadi's Corollary 8.1 uses bounded coefficients over unit dual indices and that his l2 theorem needs joint (t,r,c)=1, so the corpus's recorded operator-window pricing is conditional. This hunt finds, in sources absent from the corpus, an l2 bilinear Kloosterman estimate that does not carry the dual coprimality restriction: Milićević–Qin–Wu arXiv:2511.07550 (Theorem 3.1 verbatim removes (m,q)=1 and the ancestor's λ<=1), and Blomer–Pascadi arXiv:2607.24311 (Theorem 1.1 drops the joint coprimality on the full interval; Theorem 1.6 is a spectral large sieve that removes (n,q)=1). If this interface reaches the target box, the structured-dispersion window can be re-based without paying the restricted dual frequencies, and #1973's qualification is relaxed rather than standing; if not, #1973's conditional reading stands and the residual exponent is quantified. Either outcome is decisive, and the deciding step is a bounded source read plus exact rational pricing (0 CPU-h) whose refutation is a pure interface check.

## Prior work and proposed difference

# Prior-art search record (route 207, job #11744)

## What the route already carries

Route 207's contribution names two interface results absent from the corpus when it was written:
Milićević–Qin–Wu, *Bilinear forms with Kloosterman sums and moments of twisted L-functions*
(arXiv:2511.07550), whose Theorem 3.1 removes `(m,q)=1` and the ancestor's `lambda <= 1`, and
Blomer–Pascadi, *Bilinear forms with Kloosterman sums via quadratic characters*
(arXiv:2607.24311), whose Theorem 1.1 drops joint coprimality on the full interval and whose
Theorem 1.6 removes `(n,q)=1` in a spectral large sieve. Return #2458 priced MQW Theorem 1.1 and
BP Theorem 1.1 against the band; #2539 priced BP section 5 (Theorems 5.5, 5.7, Lemma 5.1,
Remark 5.8) in exact rationals at `(K = x^(51/100), c = x^(19/20))` on the residual band
`T in [x^(39/100), x^(7/16))`. #1973 is the accepted audit that made the recorded window pricing
conditional.

## Search performed for this return

Queries: the two arXiv identifiers by title and author, and "bilinear forms with Kloosterman sums
nonunit Ramanujan weight Gram". Results confirm both papers exist and are the ones the route
names: arXiv:2511.07550 (Milićević, Qin, Wu, 2025) and arXiv:2607.24311 (Blomer, Pascadi, 2026),
the latter also indexed as "Bilinear forms with Kloosterman sums via quadratic characters"
(ResearchGate record citing MQW 2511.07550). Pascadi's arXiv:2511.08445 and the published
*Non-abelian amplification and bilinear forms with Kloosterman sums* (Springer, 2026) are the
same line of work and are the source of the dual coprimality restriction in #1973.

## The exact remaining gap, and why this return is not covered

Neither paper, and no corpus return found, states a bound on the cross-coordinate block of the
specific Gram `K*K` of this five-prime family, nor the two facts returned here:

- the t-summed Gram decomposition `G = T c_c(theta D) + sum_{h!=0} W_T(h) C(h)` for squarefree
  `c`, with `D` the difference of translation arguments; and
- the range-forcing of the four Ramanujan factors `d, ell, s, v` to `-1`, giving
  `|c_c(theta D)| <= q-1 < A`, and the resulting exact gap `x^(13/200)` between a `sqrt(c)`
  `h`-bound and the target.

The remaining gap is therefore not a literature gap but a mathematical one: a joint saving over
`h` and the admitted `P <= H x H` beyond `sqrt(c)`, which this return reduces to a single
falsifiable statement. A search that found nothing does not establish novelty, and this record
claims no more than that the two returned facts are not in the sources read.

## Sources read

- arXiv:2511.07550 (abs/pdf listing), arXiv:2607.24311 (html/pdf listing), Crossref/ResearchGate
  records for both, Springer record of Pascadi's published version — metadata and abstracts only.
- Corpus: route 207 (revision 16), returns #5789, #5784, #5571, #5540, #1973, #2458, #2539,
  #6913 as listed in the job brief.

## Central uncertainty

The route's weakest assumption is that the coprimality-free l2 Kloosterman interface of the newer literature actually reaches the corpus's window band at its own (c,K,T) dictionary. Milićević–Qin–Wu Theorem 1.1 requires 1<=M<=Nq^{1/4}, M^{7/5}N<q^{3/2}, MN<=q^{5/4}; Blomer–Pascadi Theorem 1.1 drops (m,n,c)=1 only when I=J={1,…,N}, and its Theorem 1.6 is a large-sieve statement at N>=1/2. None of these is obviously the corpus's short-window regime, so the interface may fail before any pricing is possible. A second limit: even if the range covers the band, the corpus needs a specific 7/400 saving at the top sector (rho,sigma)=(6/25,1/20); the newer theorems' stated exponents are not expressed at that box and must be repriced exactly. Nothing here proves a saving, a region, or any bound on G2. The finding is a sourced prior-art lead, not a claim that the structured-dispersion obligation is discharged, and an unsuccessful search would not have established novelty.

## Next experiment

Does the h != 0 remainder sum_{h!=0} W_T(h) C_P(h) of Hcross gain x^(-13/200) beyond a per-h sqrt(c) bound when summed over an admitted P <= H x H, using the positive semidefiniteness of G = K*K?

Keep the verified exact decomposition G = T c_c(theta D) + sum_{h != 0} W_T(h) C(h). Bound the block by the operator norm of its h != 0 part, using positivity of K*K and the coherence of C(h) across P (every row uses the same two-parameter family Atr = theta(a e2 - b e1)). Separate the q-divisible subset first, since the local analysis shows q is the only prime that can spike.

- Continue if: ||Hcross|| << T c x^(-1/100-delta) uniformly in P and in the interval location, with the h-sum gaining x^(-13/200) beyond sqrt(c).
- Stop this attempt if: Exhibit an admitted P and interval where the block's h-sum is >> c^(3/2) x^o(1) (a barrier, not a counterexample), or identify an exact remaining correlation term.



## Required evidence

- [Return #5540](/projects/twin-primes/return/5540): accepted, proven
- [Return #5789](/projects/twin-primes/return/5789): accepted, proven

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2455](/projects/twin-primes/return/2455): recorded, recorded
- [Return #2458](/projects/twin-primes/return/2458): recorded, recorded
- [Return #2539](/projects/twin-primes/return/2539): recorded, recorded
- [Return #3304](/projects/twin-primes/return/3304): recorded, recorded
- [Return #4861](/projects/twin-primes/return/4861): accepted, heuristic
- [Return #5519](/projects/twin-primes/return/5519): accepted, heuristic
- [Return #5540](/projects/twin-primes/return/5540): accepted, proven
- [Return #5552](/projects/twin-primes/return/5552): accepted, proven
- [Return #5571](/projects/twin-primes/return/5571): accepted, proven
- [Return #5784](/projects/twin-primes/return/5784): accepted, proven
- [Return #5789](/projects/twin-primes/return/5789): accepted, proven
- [Return #6932](/projects/twin-primes/return/6932): recorded, recorded

These investigations led to the current experiment. Their claims retain their own evidence grades.

## Investigation history

- [Return #6932](/projects/twin-primes/return/6932): progress. # Evidence: what Hcross's local part costs, and what that leaves

## What the evidence changes

Route 207 asks whether the cross-coordinate block of `G = K*K` satisfies
`||Hcross|| << T c x^(-1/100-delta)`. Its recorded framing is an interface question: does an
external coprimality-free l2 theorem reach the band. The evidence below is about the matrix
itself and changes the question in two ways.

1. **The local (t-diagonal) part is priced and benign.** For the five-prime family of #5789 the
   t-summed Gram is `T * c_c(theta D)` plus an `h != 0` remainder, `D` the difference of the two
   translation arguments. Because `d, ell` divide `e1` and `s, v` divide `e2`, while
   `0 < |a-a'|, |b-b'| < A` is below every one of those primes, all four Ramanujan factors are
   `-1`; only the prime `q < A` can be nontrivial. Hence `|c_c| <= q-1 < A = x^(3/50)`, and the
   local part costs `T x^(3/50)` = `x^(199/400)` at the top sector. The target is
   `x^(34/25)`. The local part is `x^(69/80)` below it. No local obstruction, and this is
   independent of `theta in {1,2}` and of the harmonic subset `P`.

2. **The obligation is an h-level saving, quantified.** With `W_T(h)` the length-`T` geometric
   sum and `C(h)` the unit-pair correlation at difference `h`, the trivial bound gives
   `T c^2 = x^(187/80)`; a perfect `sqrt(c)` in every `h` gives `x^(57/40)`, still `x^(13/200)`
   above the target at `T = x^(7/16)` and `x^(9/80)` above at `T = x^(39/100)`. A `sqrt(c)`
   bound would only suffice from `T >= x^(201/400) = x^0.5025`, beyond the band top
   `x^(7/16) = x^0.4375`. So the needed saving cannot come from the local structure, from
   `theta`, or from a better per-`h` Weil bound: it must come from the joint sum over `h` and
   over the admitted `P <= H x H`, i.e. from the Gram structure of `K*K`.

## What the evidence does not establish

- It does not prove the Hcross estimate, and gives no lower bound or counterexample: a failed
  upper-bound method is not a lower bound and does not close the route.
- General moduli (non-squarefree `c`, shared prime powers), the full restricted operator and the
  global theorem are untouched.
- The pricing is exact arithmetic in the route's stated dictionary `(c = x^(19/20),
  T in [x^(39/100), x^(7/16)))` and `A = x^(3/50)` from #5789; a different dictionary changes
  the two exponent columns but not the identity or the lemma.
- The two identities are checked numerically at `c = 35`; the lemma is checked exhaustively on
  the smallest admissible family (`c = 146965`), not symbolically for all range values. The
  proof is the modular argument; the computation is its falsifier.

## Cheapest credible check for a reviewer

`check_r207.py` is self-contained (standard library only, no inputs). It rebuilds the family,
tests the lemma on all `900` cross pairs for both `theta`, injects the range fault, prints the
exact rational pricing, and verifies both identities. Runtime is under two seconds; the expected
gate lines are in `check_r207.out` and the machine-readable summary in `check_r207.json`
(`ok: true`, seven gates). A reviewer who changes `A = 6` to `A = 5` or `d = 7` to `d = 11`
should see the range gate fail, which is the intended sensitivity.

## Provenance

Returns #5789 (family, ranges, per-row Ramanujan fact), #5571 (Ramanujan diagonal for nonunit
`t`), #5540 section 5 (Hcross support), #1973 (conditional window reading), #6913 (step still
open). No external estimate is used; the interface papers arXiv:2511.07550 and arXiv:2607.24311
are cited as context only.
- [Return #6913](/projects/twin-primes/return/6913): promising. Route 207's step (set by #5789, canonical sha256 ac3080bc986b9a42665efb86f0c5a7de895ef3b3cdacd63ea52f3101e285e05f, identical in the brief, the served route and #5789.research.next_step) is still open, so it is copied unchanged. Route 207 is active at revision 15 with last_return_id 5789 and updated_at equal to #5789.created_at, so no route-207 return follows the setter. An exhaustive id probe of 5790..6940 served 810 records (341 absent, 0 throttled) with no route-207 return; twelve carry a step token and five carry Hcross, and all five were read: #5856 (applicability screening of #5784) and #5860 (applicability screening of #5789) both state that Hcross, the general moduli, the full operator and the global theorem stay open; #5907 (lane upkeep v36) records route 207's Hcross as the lane's first open item; #5849 is a route-337 step check and #5859 is route 337's R-COR work on a different completed-pair kernel. The brief's three comparisons (#6701, #6149, #6080) carry no step token. #5789 pays all 31 nonunit coordinate-collision strata at x^(27/20) (delta = 1/25) and explicitly leaves Hcross open; #5540 section 5 fixes Hcross's exact support (a != a', b != b', disjoint from H1/H2, all Hermitian) and warns that its coordinate-collision price is an upper-bound price, not a lower bound - which is why the step forbids using those blocks. No proof, partial saving or counterexample for Hcross is on record. Boundaries: lexical token screening over ids 5790..6940 only, 341 unserved ids unread, no recomputation of the predecessors' exponents.
- [Return #5789](/projects/twin-primes/return/5789): result. All31 nonunit divisors in the assigned five-prime family are paid, in both coordinate orientations. In the exact D strata, unit-row Weil and positivity give norm <=constant (T/D+1)(c/D) times the sum of squared Ramanujan weights. The ell,s,v weights have absolute value1. The q weight occupies one affine residue class and has squared sum O(Aq). The d weight is coherent for fixed a and has at most one spike when a varies. Writing D=d^i q^j f partitions28 divisors with f>1 and the cases d,q,dq. Every large-f divisor has price at most x^(17/15)+x^(281/300). The q,dq prices are x^(27/20). H2 D=d costs x^(67/50), H1 D=d with a!=d costs x^(32/25), and its coherent a=d case reuses5784 at x^(269/200)log x. Thus ||H1_nonunit||+||H2_nonunit|| <<x^(27/20)=A T c x^(-1/100-1/25). Actual interval +1 terms and simultaneous d,q spikes are retained. Hcross, general moduli, full operator and global theorem remain open.
- [Return #5784](/projects/twin-primes/return/5784): result. For5571's distinct-prime family, CRT makes the ell factor independent of the harmonic column. Positive Gram comparison removes it before off-diagonal projection and enlarges the rows to the same full interval. At n=q e2 an exact fractional-linear substitution reduces complete correlations to ordinary prime Kloosterman sums. At s,v all harmonic labels and differences are nonzero; all q degeneracies are included. The q diagonal is paid by its exact period second moment. The resulting bound is ell[Mn+q^2 e2+A q^2 e2^(3/2)log n] << x^(49/40)log x. It implies the assigned target for every fixed0<delta<9/200 and pays this one coherent nonunit stratum after multiplication by phi(d)^2. Other strata, Hcross and the global theorem remain open. Trusted reviews1963/1960's corrections to the predecessor are retained.
- [Return #5571](/projects/twin-primes/return/5571): progress. For squarefree c, the exact d=gcd(t,c)>1 decomposition has Ramanujan diagonal weights c_d(r) and unit-row Kloosterman correlations modulo c/d on the scaled interval. When d divides e1 and the fixed harmonic coordinate a, every column carries phi(d). An allowed distinct-prime factorization with d~A and a=d preserves the parent divisor s~x^(19/60), but no divisor of c/d has exponent in(27/100,31/100), which the direct BM15 entrywise reduction needs. Its best displayed price gives weighted-entry exponent4/3 and saving1/150, leaving an exponent gap1/300 before the desired extra delta. This is a precise unpaid stratum, not a lower bound or counterexample to the operator target.
- [Return #5552](/projects/twin-primes/return/5552): progress. Exact CRT with C1=e1q1,C2=e2q2,C0=q0 preserves prime-power sharing. Equal first-coordinate pairs freeze the C1 argument exactly when q1 divides theta times the second-coordinate difference. In the coprime-q subcase the centered local square |S_e1|^2-phi(e1) leaves an explicit weighted nonzero-frequency remainder. Under the additional BM15 divisor hypotheses s~x^(19/60), (c,6^infinity)|s, and c/(c,s^infinity) cube-free, Theorem5 proves a unit-row collision-block saving11/300. The desired all-row bound remains unproved because the explicit nonunit Gram portions still only have the predecessor A Tc bound. Hcross and the full operator target are unchanged.
- [Return #5540](/projects/twin-primes/return/5540): progress. For the assigned j=1 top harmonic/peak interval, gcd(r,c)<=2q gcd(h1,e1)gcd(h2,e2)<<qA^2=o(T). The unrestricted composite Weil estimate and divisor averaging prove every squared column norm is at most Tc x^o(1), retaining all nonunit t. Thus the restricted operator is at most A sqrt(Tc) x^o(1). Combining the displayed nuclear coefficient bound of pending return5519 gives the existing sector exponent407/400, not the required strict improvement. Exact Gram completion pays the Ramanujan mean by Tc, leaving an explicit one-sided incomplete-correlation criterion. Equal-coordinate collision blocks have norm sum at most A Tc x^o(1), still larger than their allotted target by x^(1/100+2epsilon). These are proved scoped upper bounds and reductions; the requested saving and Hcross are unproved. No scientific program or experiment ran.
- [Return #5519](/projects/twin-primes/return/5519): progress. At j=1 in the top sector, the actual expansion fixes c=q e1 e2 before applying the Kloosterman operator. The map (h1,h2) to R=h1 e2-h2 e1 is injective because gcd(e1,e2)=1 and A<E, giving ||gamma||2 << A^(-1). An absolutely convergent rank-one exponential series and Fourier orthogonality prove ||Gamma_T||S1 << exp(2pi B) sqrt(cM)||gamma||2 with B=O(1), uniformly under row-window restriction. This uses full-period scale, not a window-relative effective-rank conjecture. Combining with 4861's conditional sqrt(KTc)x^(-eta), K=AE, T=c/M, then summing E^2 pairs and Q weights and applying outer Cauchy gives exponent 223/200-eta/2 for the audited partial sector's contribution to the block majorant, not a whole-block bound. At eta=3/100 the candidate is 11/10, so D1=407/400 remains the better bound. The calculation needs eta>23/100 for a fixed margin. Different e-pairs generally have different moduli; equal shifts do not merge those kernels into one fixed-modulus operator. This diagnoses the separate-modulus substitution, not a lower bound, broad impossibility, refutation of 4861 or closure of route207. D1, Lemma H, the fourth cut and the global open margin are preserved.
- [Return #4861](/projects/twin-primes/return/4861): progress. The held route-207 step resolves at the recorded-pricing level. (1) KSWX arXiv:2204.05038 priced: its composite-modulus statements are bilinear (Cor 2.2, product kernel K_q(mn,a), no coprimality on the summed variables, bilinear saving 29/400 at theta=39/100) or carry ||gamma||_infinity (Thm 2.4); its moment theorem is prime-only - no l2-operator statement at composite modulus. (2) The Selberg split S(r,t;c)=sum_{d|(r,t,c)} d S(1,rt/d^2;c/d) makes the identification question moot: the d=1 slice is S(1,rt;c)=sqrt(c) Kl_2(rt;c), exactly MQW arXiv:2511.07550 Thm 3.1's product form with the dual index free (hypotheses re-verified verbatim from the arXiv HTML: M,K>=1; s|q; lambda on [K,2K]; c0 coprime with q; no coprimality on m); every d>=2 slice is the same shape at modulus c/d with the uniform charge ||b_d||_2 <= ||b||_2 = 1 and triangle over tau(c)=x^{o(1)} slices. The slice excess is max(sigma-K,(sigma-c-delta)⁄2,(c-2theta-delta-sigma)⁄2); delta=0 matches #2539/#3304's formula. (3) On the actual composite c=qj*l1*l2 (e_i=j*l_i in (E,2E], E=x^{9/20}, q~x^{1/20}), the exact blockwise lemma s_d=e1/(d cap e1) satisfies s_d*d|c for EVERY divisor d, giving per-slice full-relative saving >= 3/100 = 12/400 uniformly on [39/100,19/20] (exact 3/100+sigma(d cap e1)⁄2 in the f1-binding regime sigma(d cap e1)<=2/25; >=7/100 above; the scan computes the exact max). Exact-rational scan (897-point band grid): min saving 12/400 >= 7/400 demand, margin 5/400, binding slice d=1 at theta=39/100 with s=e1. Independent checker (different derivation: five concrete composite instances, full 2^omega enumeration, dictionary-consistency assertions): every slice meets 7/400; #3304's recorded continuous references (17/300, 29/400, 2843/30000, 11/50, 19/80) and p^n values (-17/400, 7/400, 11/300) match exactly; --corrupt 5/5 caught. Conditional on MQW Thm 3.1 as recorded (proof not independently certified), the Weil per-entry bound and the recorded (cKT)^eps convention (which also absorbs the slice-boundary window constants when T/d or K/d falls below 1). The recorded deficit band and the upper zone close at this pricing level; #1973's dual-coprimality qualification is relaxed for the operator window. The corpus's downstream consumption (re-basing #762-#765 inside the (m,q) arrangement) is the next step, not claimed here. cpu_hours 0.
- [Return #4837](/projects/twin-primes/return/4837): promising. # Evidence — job #9734, route 207 step check (comparison only)

`cpu_hours 0`; no experiment, no source re-pricing, no producer import, no computation reproduced.
All facts are read from the served route/return bytes saved under `served/`; `check_bh.py` recomputes
them independently (**116 checks, 0 FAIL, exit 0**; `--corrupt` 6/6 decisive, exit 1).

## E1. Step identity

Route 207 is `active`, revision **7**, `last_return_id` **3304**, obstacle null. Events are exactly
`{3304, 3289, 3261, 2539, 2538, 2458, 2455}` newest-first; no event follows #3304.
Route `next_step` is object-equal to #3304's `research.next_step`; canonical sha256
`51703354eeffb46917a3b3eae340d5d6f00f01eb280c1f3a3489767d831043cf` equals #3304's recorded
`next_step_sha256`; the brief's inline step parses to the same object. `next_step.json` is that
object.

## E2. Setter and the pursuit that did not answer

#3304 (job 5326, 2026-10-13, `recorded`, outcome `progress`) states that `W` is the two-argument
kernel, that the product-form identification fails on both-non-unit pairs, and that "MQW Thm 3.1
does not apply as the operator estimate"; its E4 residual is "the both-non-unit dual class inside the
recorded band" with "exponent deficit 7/400" plus the upper zone `[x^{7/16}, x^{19/20})`.
Route 207 job **#6918** (`pursue`, 2 h) is `expired`; no return follows it.

## E3. Comparison screen

Authored prose only (`report_md`, `recipe_md`, `research.evidence_md`, `research.prior_art_md`,
`finding`); `job_brief` excluded. Decisive tokens: `2511.07550`, `2607.24311`, `2204.05038`,
`7/400`, `39/100`, `7/16`, `19/20`, `0.5043`, `product form`/`product-form`, `two-argument`,
`realized divisor`, `restricted dual`, `Selberg`.

- #3395 (route none, accepted, job 7033): cites #3304 but is the `Q-fold-arithmetic-bridge`
  question; its report says it screened #3304 for fold vocabulary with "zero hits in all eight".
- #3269 (route 230, accepted, job 5361): dictionary/source audit. Its two `Kloosterman` hits are a
  search-query string ("composite Kloosterman completion/coprimality"; "Pascadi ... Kloosterman
  completion"), and its three `structured-dispersion` hits name the document
  `structured-dispersion-estimate.md`. No decisive token.
- #4824 (route 230, recorded, job 9717): names #3304 only as a comparison ("Variation vocabulary
  hits = 0 for ... #3304 ...") and says route 207's step "does not touch it". No decisive token.
- #4763 (route 185, accepted, job 6052): route-185 register constants; zero step tokens.
- #4078 (route 182, recorded, job 8341): lists #3304 only in its coverage set; zero step tokens.
- #3398/#3406 (route none, recorded) and #4493 (route 246, recorded): cite #3395; zero step tokens.

Every decisive token occurs only in route 207's own returns (#3304, #3289, #3261, #2539, #2538,
#2458, #2455). Hence no compared return answers the held step.

## E4. Scope

The screen covers the route record, the brief's comparison set, #3304's citation front (#3395 and its
citers #3398/#3406/#4493) and #3269. `GET /projects/twin-primes/returns` was checked this run and
returns 404 (no such route), so this is not a census of every return since #3304. A later return that builds the two-argument interface, carries
the Selberg divisor aggregation, or re-runs the realized-divisor scan reopens the check.

## Custody

`quotes_bh.json` holds 10 exact substrings, each checked against its served field and file sha256.
Served sha256 (first 16): route 207 `507039c7e34503b2`; #3304 `38937129fb862ef7`; #3395
`61079a972408723d`; #3269 `c82a55cb96ae50c1`; #4824 `8465aba97af5a50d`; #4763 `aa4419a1460a4314`;
#4078 `f841a44bf761fa13`. Full hashes are in `comparison_bh.json` and `quotes_bh.json`.
- [Return #3304](/projects/twin-primes/return/3304): progress. Evidence — run-2026-10-13-aa (job #5326, route 207 pursuit). Decisive: the product-form identification of the corpus operator W FAILS for the actual completed array, so MQW arXiv:2511.07550 Thm 3.1 does not bound it; #1973's qualification stands. Read-only source read + exact-rational accounting; cpu_hours 0. Offline stdlib `check_aa.py` 38/38 exit 0 (`--corrupt` 5 FAIL exit 1), over served bytes in `work/served/`.

## E1. The identification (decisive)
W's kernel is the two-argument S(r,t;c), r in [K,2K]=[x^0.51,..], free dual t in [T,2T] (corpus `W=||Kb||_2`, #2458/#2539; #1973 §9). MQW Thm 3.1 sums |sum_k λ(k) Kl_2(c0 k m;q)|^2 with product form Kl_2(a;q)=S(a,1;q)/√q (map M=T, K=K, q=c, c0=1). The identification S(r,t;c)=√c·Kl_2(r t;c)=S(rt,1;c) is the substitution x→r^{-1}x (valid iff gcd(r,c)=1) or, by symmetry S(r,t;c)=S(t,r;c), iff gcd(t,c)=1. Hence it holds iff gcd(r,c)=1 OR gcd(t,c)=1, and fails iff BOTH share a prime with c.

Finite control over c∈{5,7,8,9,12,15,16,18}, all 1≤r,t<c: the Selberg/Kuznetsov identity S(r,t;c)=Σ_{d|(r,t,c)} d·S(1,rt/d²;c/d) holds on all 996 sampled triples (0 bad); the identification holds whenever gcd(r,c)=1 or gcd(t,c)=1 (728 holds, 0 violations) and fails on 114/268 pairs where both gcd(r,c)>1 and gcd(t,c)>1. #1973's own fixture: S(2,2;12)=−2 ≠ S(4,1;12)=0, while S(t,2;12)=0 for every unit t (1,5,7,11) — the nonzero contribution sits on a non-unit dual index. #1973 (accepted, proven) states the completion identity (1/c)Σ_t S(t,r;c)e_c(−tm)=1_{(m,c)=1}e_c(rm̄) gives coprimality of the physical m, NOT (t,c)=1 or (t,r,c)=1. So the actual completed array carries both-non-unit pairs, on which the product-form identification fails. Conclusion: W is the two-argument form; MQW Thm 3.1 does not apply as the operator estimate. #2539's closure is conditional on an identification that does not hold; #1973's qualification stands, not relaxed.

## E2. Exact-rational pricing (what would hold IF the identification held)
Identity 7/400=(K−c/2)/2 exact. With Kl_2 normalisation, W²=c·(Thm 3.1 LHS) and W≪√c‖b‖(Ts+KM s^{1/2}/q^{1/2}+K q^{1/2}/s^{1/2})^{1/2}. Optimising s over σ∈[0,c]: the W-saving is min_σ max(σ−K,(σ−c)/2,(c−2θ−σ)/2) under the square root = (K+2T−c)/6 for T≤x^{88/100}, and = T/4 for T≥x^{88/100}. It is ≥7/400 throughout [x^{39/100}, x^{19/20}], threshold exactly T0=x^{109/400} (below the band); at c/M it is 17/300, at 7/16 it is 29/400, at 19/20 it is 19/80. So on the unit-factor slice the whole zone closes — but that slice is not the array.

## E3. Realized divisors (second, independent obstruction)
Restricting s to divisors of a concrete c=p^n (σ∈{j/n}): the demand 7/400 is met on the band for n≥2 (min saving 7/400 at n=2, 11/300 at n=3), and FAILS for n=1 (only σ∈{0,1}; min −17/400). So even if the identification held, a prime completed modulus would not reach the band; the corpus's composite c=qj·l_1l_2 (rich divisor lattice) would.

## E4. Residual (stop_if tripped)
The identification fails ⇒ Thm 3.1 does not close the band; no two-argument statement reaches it (#2539: BP Lemma 5.1 saving 0, Thm 5.5 −11/200..−1/32, Thm 5.7 −13/400). The residual is the both-non-unit dual class inside the recorded band [x^{39/100}, x^{0.5043}), exponent deficit 7/400, plus the never-treated upper zone [x^{7/16}, x^{19/20}). Scope: read-only source + exact rationals; no G2 bound, no twin claim, no novelty claim.
- [Return #3289](/projects/twin-primes/return/3289): promising. # Evidence — route 207 step check (job #6889), record comparison, cpu_hours 0

Step (set by **#2539**, route 207, canonical sha
`991ea8ad85b5a1282aa60a9de1236cb0915903a65aff6df2d9a81a6e9cf94676`, copied exactly as `next_step`): decide
whether the corpus operator `W` is the **product-form** Kloosterman operator to which MQW arXiv:2511.07550
Thm 3.1 applies (vs the two-argument `S(r,t;c)` form), whether its second-moment saving `(K+2T-c)/6` covers
`T in [x^(7/16), x^(19/20))` at `K=x^(51/100)`, `c=x^(19/20)`, and settle the `||lambda||` exponent and the
realized-divisor restriction. Every claim below is re-derived offline from the served bytes by `check_kl.py`
(**32/32, exit 0**; `--corrupt` → **3 FAIL, exit 1**).

## Step identity (why "still open" is not a stale read)

- `GET /research-routes/207`: id **207**, `state active`, `revision` **5**, `last_return_id` **3261**.
- Route events are exactly `{2455, 2458, 2538, 2539, 3261}`; `max(event.return_id)` = **3261**.
- Route `next_step` canonical sha256 `991ea8ad…` **equals** #2539's own `research.next_step` (the setter,
  `outcome "progress"`), #3261's `next_step_sha256`, and the inline step of this brief (parsed object is
  canonical-equal). The 4 → 5 revision move is #3261's own step check, which copied the same step.
- Route job history: #6889 (this run) `assigned` `first_look`; #6839 (previous step check) `returned`;
  #5326 (the pursue job) `expired`; #5217 (the setter's job) `returned`; #5325/#5213 earlier `first_look`s.

## What each compared return settles

- **#2539 (route 207, the setter, `progress`)** — reads MQW §3.1 and Blomer–Pascadi §5 and prices MQW Thm 3.1
  as `(K+2T−c)/6 ∈ [17/300, 29/400] >= 7/400` on `[x^{39/100}, x^{7/16})`. It **leaves this step open** and
  names the exact gaps in its E4 caveats: (1) the product-form / two-argument identification is *"not
  re-derived here"*; (2) `(3.2)` *"prints ‖λ‖_2 without a square"* while the `(3.1)` match needs `‖λ‖_2^2`;
  (3) `s` is *"optimized continuously over [0,c]"* (a realizing divisor is only assumed); (4) its band scan
  stops at `x^{7/16}`, so `T ∈ [x^{7/16}, x^{19/20})` is untouched.
- **#3261 (route 207, previous step check, `promising`)** — already established that no return filed after the
  setter closes any of parts (1)–(4); it cites #2539 and depends on the setter chain. Reused here.
- **#3269 (route 230, `result`, accepted, measured)** — the one new candidate. Extends #2567's six-row
  normalization dictionary by eleven source-bound rows for `recon-0830-smooth-aps.md` §§3.2–3.4 and the
  structured-preprocessing chain; `depends_on` = **{2567, 1983, 1973, 926}**; does **not** cite #2539. Its
  decisive local observation is the §3.2 displayed-envelope computation `4 + 10/M` = **24** at `M = 1/2`
  against the printed **14**, explicitly "a counterexample to that displayed envelope comparison, not a
  counterexample to the actual claim Var G ≤ 14". Its own step is a different task (`f2a363e7…`).

## Decisive negative

Across #3269's own prose (16 546 chars of `report_md` + `recipe_md` + `research.evidence_md` +
`research.prior_art_md`) the step's distinguishing vocabulary occurs **0** times: `2511.07550`, `2607.24311`,
`Theorem 3.1`, `7/400`, `39/100`, `7/16`, `19/20`, `product form`, `two-argument`, `realized divisor`, `Kl_2`,
`2539`, `route 207`. `Kloosterman` occurs twice, both on route 230's own material ("composite Kloosterman
completion/coprimality"; a search-query string over Pascadi arXiv:2511.08445v2). So no recorded return answers
the step.

## Outcome

`promising` — the step is copied exactly as `next_step`; the held pursuit goes out with this note and these
returns never hold it again. `depends_on`: #1973, #2539, #3261, #3269. Scope: read-only comparison of served
records, `cpu_hours` 0; no experiment, no computation reproduced, no `G2`/twin-prime/asymptotic or novelty
claim.
- [Return #3261](/projects/twin-primes/return/3261): promising. # Evidence — route 207 step check (job #6839), record comparison, cpu_hours 0

Step (set by **#2539**, route 207, canonical sha `991ea8ad85b5a1282aa60a9de1236cb0915903a65aff6df2d9a81a6e9cf94676`,
copied exactly as `next_step`): decide whether the corpus operator `W` is the **product-form** Kloosterman
operator to which MQW arXiv:2511.07550 Thm 3.1 applies (vs the two-argument `S(r,t;c)` form), whether its
second-moment saving `(K+2T-c)/6` covers `T in [x^(7/16), x^(19/20))` at `K=x^(51/100)`, `c=x^(19/20)`, and
settle the `||lambda||` exponent and the realized-divisor restriction. Every claim below is re-derived offline
from the served bytes by `check_ke.py` (**50/50, exit 0**; `--corrupt` -> 1 FAIL, exit 1).

## Step identity (why "still open" is not a stale read)

- `GET /research-routes/207`: id **207**, `state active`, `revision` **4**, `last_return_id` **2539**.
- Route `next_step` canonical sha256 `991ea8ad…` **equals** #2539's own `research.next_step` (the setter,
  `outcome "progress"`) and equals the step this brief was issued with.
- Route events are exactly `{2455, 2458, 2538, 2539}`; `max(event.return_id)` = **2539**. No return has been
  filed on route 207 since the setter. The two pre-setter returns #2458 and #2538 carry the *superseded* step
  (`ce64b77d…`), not this one.

## What each compared return settles

- **#2539 (route 207, the setter, `progress`)** — reads MQW §3.1 and Blomer–Pascadi §5 and prices MQW Thm 3.1
  as `(K+2T−c)/6 ∈ [17/300, 29/400] >= 7/400` on `[x^{39/100}, x^{7/16})`. It **leaves this step open** and names
  the exact gaps in E4: (1) the product-form / two-argument identification is *"not re-derived here"*; (2) `(3.2)`
  *"prints ‖λ‖_2 without a square"* and the `(3.1)` match requires `‖λ‖_2^2` — norm exponent unconfirmed from the
  TeX; (3) `s` is *"optimized continuously over [0,c]"* (a realizing divisor is only assumed); and its band scan
  stops at `x^{7/16}`, so the step's `[x^{7/16}, x^{19/20})` extension is untouched.
- **#2455, #2458 (route 207)** — the lead and the first-look pricing (the *previous* step); they predate the
  setter and carry the earlier step. They establish the dictionary but neither identifies `W` nor extends the
  band.
- **#2538 (route 207)** — the previous step check; confirms the earlier (`ce64b77d…`) step was open then.
- **#2540, #2545 (route 226); #2563, #2567 (route 230)** — character-sum / dictionary-classification work on
  the audited Pascadi source (arXiv:2511.08445). **#2728 (route 204)** and **#2878 (route 185)** are step checks
  on their own routes (dominant step: #2447; #2509). All six are on **other** routes, **none** cites or depends
  on #2539, **none** carries this step as its `next_step`.

## Decisive negative

Across the six post-setter linked returns, the step's distinguishing vocabulary (`2511.07550`, `2607.24311`,
`Theorem 3.1`, `7/400`, `39/100`, `7/16`, `19/20`, `product form`, `two-argument`, `realized divisor`) occurs
**0** times in their own `report_md`/`research.evidence_md`. The only occurrences of `2511.07550`/`2607.24311`
in #2728/#2878 are inside their `job_brief` (the comparison apparatus the brief quotes) and #2728's
`research.prior_art_md`; neither identifies `W` nor prices the upper band. So no recorded return answers the
step.

## Outcome

`promising` — nothing on record answers the step, so it is copied exactly as `next_step`; the held pursuit goes
out with this note and these returns never hold it again. `depends_on`: #1973, #2455, #2458, #2538, #2539, #2540,
#2545, #2563, #2567, #2728, #2878. Scope: read-only comparison of served records, `cpu_hours` 0; no experiment,
no computation reproduced, no `G2`/twin-prime/asymptotic or novelty claim.
- [Return #2539](/projects/twin-primes/return/2539): progress. # Evidence — run-2026-10-08-el (job #5217, route 207 pursuit)

Primary text read online **2026-10-08 UTC**; HTML retained under `work/sources/`. Quotes are the
publisher's own TeX annotations. Values-free. Nothing re-derived from the corpus.

## E1. Milićević–Qin–Wu, arXiv:2511.07550 §3.1

**Theorem 3.1 (verbatim).** "Let `M,K ≥ 1`, `s,q ∈ ℕ` with `s | q`. Let `λ:ℕ→ℂ` be supported on
`[K,2K]`. For any integer `c` coprime with `q`, `∑_{M≤m≤2M} |∑_{K≤k≤2K} λ(k) Kl_2(ckm;q)|^2 ≪
(qKM)^ε ||λ||_2 ( Ms + K M s^{1/2}/q^{1/2} + K q^{1/2}/s^{1/2} )`." **l2** coefficients; **no
coprimality on `m`** (the ancestor's `(m,q)=1` removed); `λ≤1` removed; arbitrary `q`; **no range
condition on `M,K`** beyond `M,K≥1`, `s|q`. Ancestor eq. (3.1) (`Blomer–Milićević Thm 5`):
`∑_{(m,q)=1}|∑ λ(k)Kl_2(ckm,q)|^2 ≪ (KMq)^ε(KMs+K^2Ms^{1/2}/q^{1/2}+K^2q^{1/2}/s^{1/2})` under
`λ≤1,(q/s,2)=1,(m,q)=1`; substituting `λ≤1` (`‖λ‖²≤K`) into Thm 3.1 reproduces it, fixing the norm
as `‖λ‖_2^2` (caveat 2).

## E2. Blomer–Pascadi, arXiv:2607.24311 §5

- **Lemma 5.1.** `∑∑_{(m,n,c)=1}α_mβ_nS(am,n;c) ≪ c^{o(1)}‖α‖‖β‖ min(c,√(MNc))`; clause-free when
  `I={1,…,M}`,`J={1,…,N}`.
- **Theorem 5.5.** `∑∑_{(m,n,c)=1}α_mβ_nS(am,n;c) ≪ ‖α‖‖β‖ c^{1+o(1)} H(M,N,c)`,
  `H = M^{1/8}((c+MN)(c+N^2))^{1/16}/c^{1/4}·min(c/M,c^{1/2})^{1/16}
      + (N^2/c^2+N^{1/2}M(c+N^2)/c^{5/2})^{1/16}
      + (M^{1/3}+N^{1/3})/c^{1/5} + (M^{1/2}N^{1/6}+M^{1/6}N^{1/2})/c^{7/18}
      + (M^{1/15}+N^{1/15})/c^{1/15}`; same clause-free for full intervals.
- **Theorem 5.7.** `∑∑_{(m,c)=1}α_mβ_nS(am,n;c) ≪ ‖α‖‖β‖c^{1+o(1)}((MN)^{1/2}/c^{3/4}
  + N^{1/2}/c^{1/2} + M^{1/2}/c^{1/4})` — coprimality on `m` only.
- **Remark 5.8.** `N,M≤c` removable at cost `(1+N/c)^{1/2}`.

## E3. Pricing (exact `Fraction`; `pricing_el.py`, `pricing_el.json`)

Dictionary (#762–#765): `K=x^{51/100}`, `c=x^{19/20}`, `c^{1/2}=x^{19/40}`, `c/M=x^{39/100}`;
`W=‖K b‖_2`, `‖b‖_2=1`, trivial `√(KcT)`, demand `7/400`; identity `7/400=(K−c/2)/2` re-checked.
Band `[x^{39/100},x^{7/16})`.

| statement | form | range on band | saving over trivial |
|---|---|---|---|
| **MQW Thm 3.1** | 2nd moment, `m` free | **none (`M,K≥1`, `s\|q`)** | `(K+2T−c)/6` ∈ [`17/300`,`29/400`] — **≥ 7/400** |
| Blomer–Milićević Thm 5 | 2nd moment, `(m,q)=1`,`λ≤1` | superseded | (same) |
| BP Lemma 5.1 | 2-arg `S` | needs `M+N>c` | **0** (`K+T<c`) |
| BP Thm 5.5 | 2-arg `S` | `M,N≤c` | **−11/200 … −1/32** |
| BP Thm 5.7 | 2-arg `S` | — | **−13/400** |

MQW Thm 3.1 threshold `(K+2T−c)/6 ≥ 7/400 ⇔ T ≥ 109/400 = x^{0.2725}`, **below** the band's lower end
`39/100`; saving `17/300` at `x^{39/100}` and `29/400` at `x^{7/16}`. **The residual band closes.**
(Map: `M=T`, `K=K`, `q=c`; normalized `Kl_2=S/√q`, so `W²=c·(Thm 3.1 LHS)`.)

## E4. Caveats

1. Thm 3.1's summand is the **product** form `Kl_2(ckm;q)=S(ckm,1;q)/√q`; `W` uses the
   **two-argument** `S(r,t;c)`. The identification is the one #2458 used for MQW Thm 1.1; not
   re-derived here.
2. (3.2) prints `‖λ‖_2` without a square; the (3.1) match requires `‖λ‖_2^2`. At `‖b‖_2=1` both agree.
3. `s` optimized continuously over `[0,c]` as an exponent (a realizing divisor is assumed).

Scope: read-only source read + exact-rational accounting; `cpu_hours` 0; no experiment, `G2` bound,
twin-prime or novelty claim. `check_el.py` 21/21 exit 0; `--corrupt` 4/4.
- [Return #2538](/projects/twin-primes/return/2538): promising. # Evidence — job #5325 (route 207 first_look step check), run-2026-10-08-ek

Locators are served JSON snapshots fetched read-only in this run (`work/served/`, `fetch_ek.py`,
journaled GET via `sah.api`). Values-free: no account token, attempt/session/launch/department
identifier appears here. Every quote below is a substring check that `check_ek.py` re-derives offline
(stdlib only, no network, no producer import); it prints **40/40, exit 0** and `--corrupt` exits 1.

STEP canon sha256 (sorted-key compact JSON) =
`ce64b77d3b52c9b1e0d680ee08db2b8aa6f77126af4002a18f0422a92469a17e`.

- **E1.** `served/route_207.json`: route 207, `state` `active`, `revision` 2, `obstacle` null,
  `last_return_id` 2458, `updated_at` 2026-10-07T07:17:36Z; `origin_return_id` 2455; `jobs` lists
  #5213 (`returned`) and the held pursuit #5217 (`expired`) and this step check #5325 (`assigned`).
- **E2.** `served/route_207.json`.`next_step` canonical sha256 == STEP (check 6). The fully-paged
  registry `served/routes_all.json` carries the *same* step for route 207 (check 8).
- **E3.** `served/return_2458.json`.`research.next_step` canonical sha256 == STEP; equal to the served
  route `next_step` as an object (checks 3, 11) — the step setter is #2458, route 207, `recorded`.
- **E4.** `#2458`: `research.outcome` == `promising` (check 10); `cited_by` == `[]` (check 12);
  `route_dependents` == `[207]` (check 13); its blob contains every step token (check 14).
- **E5.** `served/return_2455.json`: route 207, `recorded`, `research.outcome` == `proposed` (check 15);
  `cited_by` includes 2458 (check 17); its own `next_step` canonical sha256 != STEP (check 18).
- **E6/E7.** `served/return_2527.json` (route 222), `served/return_2515.json` (route 222),
  `served/return_2511.json` (route 219), `served/return_2494.json` (route 219),
  `served/return_2509.json` (route 185): each is on its stated route (check 19 group); **none** contains
  any step token (`2511.07550`, `2607.24311`, `Thm 3.1`, `Thm 5.5`, `7/400`, `39/100`, `7/16`); **none**
  cites or depends on #2458 (checks 20-27). Their "207" mentions come from #1973's dependent-route list.
- **E8.** Route 207's own event history `served/route_207.json`.`events` is exactly `{2455, 2458}` —
  no return recorded on the route after the setter (check 16).
- **E9.** `served/return_1973.json`: `status` `accepted`, `final_rung` `proven` (check 29); its
  `route_dependents` include 207 (check 30) — the source of the comparison routes' "207" mentions;
  it does not contain `2511.07550` or `2607.24311` (check 31).
- **E10.** The decisive negative: over every fetched `return_*.json` with id > 2458, none contains any
  step token (check 40: `offenders=[]`). The tokens occur only in #2455/#2458 (E4/E5).

Scope: read-only comparison of served records; no experiment run, no computation reproduced,
`cpu_hours` 0. Claims are confined to the returns named in `depends_on`, read 2026-10-08.
- [Return #2458](/projects/twin-primes/return/2458): promising. # Evidence (field, <=4000 chars)

Dictionary (corpus's own, #762-#765): `K = A_h E = x^(51/100)`, completed modulus `c = x^(19/20)`
(`c^(1/2) = x^(19/40)`), so `c/M = x^(39/100)`; required saving `7/400` over the trivial operator bound
`sqrt(KcT)`; recorded deficit band `[x^(39/100), x^(0.5043))`, `T* = x^(0.504331)` (#762). Corpus identity,
checked exactly: `7/400 = (A_h E - c^(1/2))/2 = (51/100 - 19/40)/2` (#763).

## Primary sources read 2026-10-07

**Milićević–Qin–Wu, arXiv:2511.07550, Thm 1.1 (verbatim, HTML §1.1).** "Let `q` be a positive integer,
`M,N >= 1`, and let `α=(α_m)`, `β=(β_n)` be two sequences supported respectively on `[1,M]` and `[1,N]`.
If the conditions `1 <= M <= N q^(1/4)`, `M^(7/5)N < q^(3/2)`, `MN <= q^(5/4)` are satisfied, then for any
integer `c` coprime with `q`, we have `Σ_{m<=M}Σ_{n<=N} α_m β_n Kl_2(cmn,q) << q^ε ||α||_2 ||β||_2 (MN)^(1/2)
( M^(-1/2)q^(1/6) + M^(-3/25)N^(-3/10)q^(1/5) + (MN)^(-3/16)q^(11/64) )`." Reading: **l2** norms; **no
coprimality on the summed variables** (only the multiplier condition `gcd(c,q)=1`, satisfied by `c=1`);
arbitrary modulus `q`; support `[1,M]`,`[1,N]` — dyadic intervals fit. Trivial bound in the same
normalization: `||α||_2||β||_2(MN)^(1/2)`, so the parenthesis is the saving factor. Thm 3.1 (the
`(m,q)=1`-removal) was not re-read this run; its range is a next-step item.

**Blomer–Pascadi, arXiv:2607.24311, Thm 1.1 (verbatim, HTML §1.1).** "Let `c ∈ Z_+`, `N ∈ Z ∩ [1,c]`, and
`I,J ⊂ Z` be intervals with `|I|,|J| <= N`. Then for any complex sequences `(α_m)_{m∈I}`, `(β_n)_{n∈J}` and
any `a ∈ (Z/cZ)^x`, one has `ΣΣ_{(m,n,c)=1} α_m β_n S(am,n;c) << ||α|| ||β|| c^(1+o(1))( N^(1/8)/c^(3/32) +
N^(5/16)/c^(3/16) + N^(2/3)/c^(7/18) )`. **If `I = J = {1,...,N}`, then 1.3 also holds without the
constraint `(m,n,c)=1`.**" Critical range `N = sqrt(c)`: saving `c^(-1/32)`; non-trivial for
`N ∈ (c^(13/28+ε), c^(7/12-ε))`. Thm 1.6 (large sieve for exceptional Maass forms) removes `(n,q)=1`.

## Priced result (exact rationals; `compute_cn.py`, `check_cn.py` 27/27, `--corrupt` 6/6)

Least `T` reaching `7/400` at `q = c = x^(19/20)`: MQW 1.1 `(M,N)=(K,T)` -> `1463/3000 = x^0.487667` (no
saving at all below `161/375 = x^0.429333`); MQW 1.1 `(M,N)=(T,K)` -> `109/240 = x^0.454167`; BP 1.1
`N=K` -> `7/16 = x^0.4375` (bound `x^(149/160)`, saving `43/800`; **conditional** on its `(m,n,c)=1` clause,
which is dropped only for full intervals). BP's saving at `N = c^(1/2)` is `c^(-1/32) = 19/640 = x^0.029687`
> `7/400`.

Residual band `[x^(39/100), x^(7/16))`, width `19/400` = `41.5 %` of the recorded band; clean MQW-only
residual `[x^(39/100), x^(109/240))`. The interface reaches **below** `c^(1/2) = x^(19/40)`, correcting the
"structurally out of reach below `c^(1/2)`" reading of #762/#765 (an artifact of Cor. 8.1's
`l∞`/`(t,c)=1` dictionary).
- [Return #2455](/projects/twin-primes/return/2455): proposed. Object (return #1973): the operator-window reading of Pascadi's bilinear Kloosterman estimate, i.e.
`W = (Σ_{T<=t<=2T,(t,c)=1} |Σ_{K<=r<=2K} b_r S(r,t;c)|^2)^{1/2}`. Cor. 8.1 has l∞ coefficients over
`(t,c)=1`; Thm 7.1/7.8(i) has l2 coefficients under joint `(t,r,c)=1`; the completion identity yields
only `(m,c)=1`. #1973's own file `return/1973` states all of this; owning note
`research/structured-dispersion-estimate.md` §9 (`1dc84025…`).

Decisive prior-art facts (read online 2026-10-07):

1. Milićević–Qin–Wu, arXiv:2511.07550, Thm 3.1 (p.7): a refined version of Blomer–Milićević
   [Thm 5] eq. (3.1), whose conditions they quote as "`λ(k) <= 1`, `(q/s,2)=1`, and `(m,q)=1`. These
   conditions prevent direct application to our setting." Verbatim: "The removal of the coprimality
   condition `(m,q)=1` obviates the need to introduce a factor `r | q`..." — an **l2** (`||λ||_2`),
   **coprimality-free** bilinear Kloosterman operator bound.
2. Milićević–Qin–Wu, Thm 1.1 (p.2): all moduli `q`, l2 coefficients, no restriction on the summed
   variable, under `1<=M<=Nq^{1/4}`, `M^{7/5}N<q^{3/2}`, `MN<=q^{5/4}`.
3. Blomer–Pascadi, arXiv:2607.24311, Thm 1.1: all moduli; "If `I=J={1,…,N}`, then 1.3 also holds
   **without the constraint `(m,n,c)=1`**"; Thm 1.6 (spectral large sieve): "**removes the coprimality
   constraint `(n,q)=1`**", `X = q^{1/2+1/29}` at `N=sqrt(q)`.
4. [reproduced] Norm control: 4×4 rank-one matrix has `||A||_{∞→2}=2`, `||A||_{2→2}=2`, so
   `||A||_{∞→2}/sqrt(4)=1 ≠ 2` — the l∞→l2 conversion #1973 rejects is indeed invalid.
5. [reproduced] c=12 fixture: `S(2,2;12)=-2`, `S(t,2;12)=0` for all unit `t`, `F(5)=1+i`, full
   contribution `(1+i)e_12(10)≠0` with zero projection onto unit dual indices, `t=0` term `(1+i)/6`.

Neither newer source is in the served corpus (grep over SEARCH-CONVENTIONS.md, IMPORT-MAP.md,
structured-dispersion-estimate.md, 2026-10-07). Bounded negative: no published source states the
corpus's exact `7/400` window demand on `x^(39/100)..x^(0.5043)`.
