erdos-problems 0.3.1 → 0.3.3

package/README.md

@@ -1,6 +1,6 @@
 # erdos-problems
 
-Maintained by Fractal Research Group (`frg.earth`).
+Maintained by SproutSeeds. Research stewardship: Fractal Research Group ([frg.earth](https://frg.earth)).
 
 > CLI and workspace for Paul Erdos problems.
 >

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "erdos-problems",
-  "version": "0.3.1",
+  "version": "0.3.3",
   "description": "CLI atlas and workspace tools for Paul Erdos problems.",
   "type": "module",
   "bin": {

package/packs/number-theory/README.md

@@ -5,6 +5,7 @@ Light starter workspace for number-theory problems in `erdos-problems`.
 Current seeded pack problems:
 - `1`: open starter workspace around a distinct-subset-sum lower-bound route
 - `2`: counterexample/archive workspace for a disproved covering-systems problem
+- `848`: decidable finite-check workspace around explicit threshold extraction and bounded closure
 
 Useful commands:
 - `erdos number-theory status 1`

package/packs/number-theory/problems/848/BOUNDED_VERIFICATION_PLAN.md

@@ -0,0 +1,43 @@
+# Problem 848 Bounded Verification Plan
+
+This note chooses the next closure lane for Problem `848`.
+
+Chosen lane:
+- bounded finite verification under the best trusted explicit threshold currently available
+
+Reason for this choice:
+- `848` is already a decidable-gap problem, not a fresh asymptotic frontier
+- the imported public thread currently reaches `N0 = 2.64 x 10^17` on 2026-03-23
+- lowering `N0` further is still valuable, but only because it reduces the finite remainder
+- a structured verification program is closer to the actual finish line than another isolated
+  threshold race
+
+Scope of this lane:
+- do not claim full closure yet
+- do not silently adopt imported threshold claims as canonical repo truth
+- instead, build the machinery that would let the repo:
+  - state exactly what range is covered by what method
+  - preserve certificates and checkpoints
+  - audit imported verification claims before relying on them
+
+Immediate bounded-verification objectives:
+1. Freeze the best imported threshold currently trusted enough to size the finite remainder.
+2. Split the finite remainder into verification regimes rather than treating it as one giant
+   interval.
+3. Define what a reproducible verification certificate must contain before the repo counts a
+   range as covered.
+4. Record prior public verification attempts, especially where external reviewers raised
+   correctness concerns.
+5. Keep the verification lane modular enough that imported proofs, local proofs, and future
+   compute runs can all plug into the same certificate surface.
+
+What success would look like:
+- the repo can say exactly which interval is covered
+- each covered interval points to a concrete certificate type and reproduction story
+- imported public verification work is either accepted with an audit note or left external
+- the remaining uncovered range shrinks monotonically
+
+What this lane is not:
+- not brute force to `2.64 x 10^17`
+- not an automatic endorsement of every public verification claim
+- not a replacement for threshold-lowering work if a cleaner explicit `N0` emerges

package/packs/number-theory/problems/848/BRANCH_COMPARISON_LEDGER.md

@@ -0,0 +1,85 @@
+# Problem 848 Branch Comparison Ledger
+
+This note closes `N848.G1.A8`.
+
+Question:
+- does the current shared witness
+  - `T = 250`
+  - `N >= exp(1420)`
+  - `eta = 10^-4`
+
+already dominate the other public case bounds in Sawhney's proof?
+
+Answer:
+- yes, at the repo's current explicit level it does
+
+## Shared witness inputs
+
+From the existing explicit notes:
+- weakest-branch main term: `0.0376113079`
+- `A*` tail at `T = 250`: `0.0005641453`
+- two-class Lemma 2.2 tail at `T = 250`: `0.0000491054`
+- one-class prime-count term at `N >= exp(1420)`: at most `1 / (1420 - 1.1)`
+- two-class prime-count term at `N >= exp(1420)`: at most `2 / (1420 - 1.1)`
+
+These are enough to compare the other public branches without changing the witness.
+
+## Case `0.0358`
+
+Repo explicit bound at the shared witness:
+- main term about `0.0356925181`
+- visible tail terms about `0.0002943490`
+- one-class prime-count term about `0.0007047713`
+
+Total visible bound:
+- about `0.0366916384`
+
+Visible reserve to `0.04`:
+- about `0.0033083616`
+
+## Case `0.0336`
+
+Repo explicit bound at the shared witness:
+- main term about `0.0334753577`
+- visible tail terms about `0.0003311781`
+- two-class prime-count term about `0.0014095427`
+
+Total visible bound:
+- about `0.0352160784`
+
+Visible reserve to `0.04`:
+- about `0.0047839216`
+
+## Case `0.0294`
+
+Repo explicit bound at the shared witness:
+- main term about `0.0293394076`
+- visible tail terms about `0.0000491054`
+- two-class prime-count term about `0.0014095427`
+
+Total visible bound:
+- about `0.0307980556`
+
+Visible reserve to `0.04`:
+- about `0.0092019444`
+
+## Honest comparison conclusion
+
+At the repo's current explicit level:
+- the `0.0377` branch remains the tightest branch
+- but it is already closed by the shared witness
+- the branches `0.0358`, `0.0336`, and `0.0294` all retain strictly larger visible reserve
+
+So the current witness appears to be a whole-proof witness candidate, not just a
+weakest-branch patch.
+
+## Remaining caution
+
+This is still a repo witness ledger, not yet a publication-ready explicit-threshold proof.
+
+What remains:
+- package the whole argument as a proposition-level explicit candidate
+- decide whether to state `N0 <= exp(1420)` as the current repo witness, or to keep it as a
+  claim-safe candidate pending a cleaner writeup
+
+But the route no longer points to another hidden branch obstruction.

package/packs/number-theory/problems/848/CERTIFIED_NUMERICAL_LEDGER.md

@@ -0,0 +1,88 @@
+# Problem 848 Certified Numerical Ledger
+
+This note closes `N848.G1.A11`.
+
+Purpose:
+- replace the displayed decimal inputs used in the current witness with conservative machine
+  intervals
+
+Scope:
+- this is still a repo-level certification note, not a formal proof-assistant artifact
+- the goal is to eliminate naked decimal approximations from the current witness candidate
+
+## Method
+
+Cutoff:
+- enumerate primes up to `5,000,000`
+
+For the Euler-product main terms:
+- multiply the truncated product in machine arithmetic
+- apply a conservative relative rounding envelope `n * eps / (1 - n * eps)` where `n` is the
+  number of factors and `eps = 2^-52`
+- apply a missing-tail lower bound using
+  - `sum_{p > P} 2 / p^2 <= 2 / P` for the `A*` product
+  - `sum_{p > P} 1 / p^2 <= 1 / P` for the `A7 union A18` product
+
+For the tail sums:
+- sum the prime-square reciprocals up to the same cutoff
+- add both a summation-error envelope and the missing-tail envelope `1 / P`
+
+## Certified upper bounds used by the witness
+
+### 1. `A*` main term
+
+Certified upper bound:
+- `A* main <= 0.0251591225`
+
+### 2. `A7 union A18` main term
+
+Certified upper bound:
+- `A7 union A18 main <= 0.0124525569`
+
+### 3. `A*` tail at `T = 250`
+
+Certified upper bound:
+- `A* tail <= 0.0005641454`
+
+### 4. Two-class Lemma 2.2 tail at `T = 250`
+
+Certified upper bound:
+- `Lemma 2.2 tail <= 0.0000491055`
+
+### 5. Two-class prime-count term at `N >= exp(1420)`
+
+From `LEMMA22_PRIME_COUNT_BOUND.md`:
+- `2 pi(N) / N <= 0.0014095427`
+
+## Certified weakest-branch witness
+
+Using only the certified upper bounds above, the weakest branch satisfies
+
+- `|A| / N <= 0.0396344729`
+
+for the shared witness scale
+- `N >= exp(1420)`
+- `T = 250`.
+
+So the certified visible reserve to `1/25 = 0.04` is at least
+
+- `0.04 - 0.0396344729 = 0.0003655271`.
+
+After the working choice
+- `eta = 10^-4`,
+
+the certified visible reserve is still at least
+
+- `0.0002655271`.
+
+## Honest consequence
+
+This note removes the main numerical discomfort in the current repo candidate:
+- the witness no longer depends only on display decimals copied from exploratory notes
+
+What it still does **not** provide:
+- a formal proof-assistant certificate
+- a publication-ready theorem writeup
+
+But it does mean the current route can now treat the witness as numerically hardened at the
+repo level.

package/packs/number-theory/problems/848/CHECKPOINT_TEMPLATE.md

@@ -0,0 +1,8 @@
+# Problem 848 Checkpoint Template
+
+- Active route:
+- Threshold ledger update:
+- Finite-check range clarified:
+- Formalization coverage clarified:
+- Public-status wording rechecked:
+- Next honest move:

package/packs/number-theory/problems/848/CONTEXT.md

@@ -0,0 +1,11 @@
+# Problem 848 Number Theory Pack Context
+
+- Family role: finite_check_number_theory_workspace
+- Harness profile: decidable_gap_workspace
+- Active route posture: finite_check_gap_closure
+
+This pack slice exists because Problem 848 is no longer a generic open dossier.
+The honest posture is:
+- preserve the sufficiently-large-N theorem
+- isolate the finite unresolved gap
+- keep threshold extraction, finite computation, and formalization coverage distinct

package/packs/number-theory/problems/848/EXTERNAL_VERIFICATION_LEDGER.md

@@ -0,0 +1,56 @@
+# Problem 848 External Verification Ledger
+
+This ledger records public verification-style claims that matter to the bounded finite
+verification lane.
+
+## Imported public items
+
+### Public thread post on 2026-03-16
+
+Source:
+- https://www.erdosproblems.com/forum/thread/848
+
+Claim:
+- a public post claimed a complete verification framework with computation through `10^7`
+  and an intended handoff to Sawhney above that
+
+Why it matters:
+- this is exactly the kind of bounded-verification claim the repo should eventually be able
+  to absorb, audit, or reject cleanly
+
+Public follow-up:
+- the same thread notes that the asymptotic handoff was overstated because Sawhney's theorem
+  does not start at `10^7`
+- external review also raised quality and verification concerns, including that the repo was
+  difficult to verify and likely incorrect
+
+Repo audit posture:
+- do not treat this as a covered interval
+- keep it as a cautionary example motivating stricter certificate requirements
+
+### Imported explicit-threshold timeline
+
+Source:
+- https://www.erdosproblems.com/forum/thread/848
+
+Imported values:
+- `7 x 10^17` on 2026-03-21
+- `3.3 x 10^17` on 2026-03-22
+- `2.64 x 10^17` on 2026-03-23
+
+Why it matters:
+- these claims change the size of the remaining finite gap
+- they do not by themselves verify any bounded interval below the threshold
+
+Repo audit posture:
+- operationally relevant
+- not yet promoted to repo-owned theorem truth
+
+## Current repo rule
+
+Imported verification work is welcome, but it only counts toward canonical coverage after the
+repo can answer:
+- what exact interval is covered
+- by what method class
+- with what reproduction path
+- and whether external criticism has been answered

package/packs/number-theory/problems/848/EXTRACTION_CHECKLIST.md

@@ -0,0 +1,114 @@
+# Problem 848 Extraction Checklist
+
+This checklist turns the current threshold ledger into the next explicit route task.
+
+Goal:
+- identify exactly what must be made numerical before the existential `N0` in Sawhney's note becomes a usable explicit threshold
+
+## Margin ledger from the public note
+
+Target extremal density:
+- `1/25 = 0.04`
+
+Rounded case bounds recorded in the note:
+- Case 1: `0.0377`
+- Case 2: `0.0358`
+- Case 3: `0.0336`
+- Final mixed-class case: `0.0294`
+
+Current margin to the target:
+- Case 1 slack: `0.0400 - 0.0377 = 0.0023`
+- Case 2 slack: `0.0400 - 0.0358 = 0.0042`
+- Case 3 slack: `0.0400 - 0.0336 = 0.0064`
+- Final mixed-class slack: `0.0400 - 0.0294 = 0.0106`
+
+Immediate implication:
+- the weakest case is the `0.0377` branch
+- any fully explicit threshold extraction has to force the accumulated error terms in that branch below the available `0.0023 N` slack
+- before spending that full slack, the repo should freeze the exact main-term constants behind `0.0252` and `0.0125`, since the public note only records rounded decimals
+
+## Proof components to quantify
+
+### A. Lemma 2.1 tail and inclusion-exclusion error
+
+Need:
+- explicit control of the prime-square tail `sum_{T <= p <= N^(1/2)} N/p^2`
+- explicit control of the small-prime inclusion-exclusion remainder
+- an explicit version of the `N^(o(1))` term coming from the small-prime product
+
+Route question:
+- can these be made comfortably smaller than the smallest slack without overcomplicating the rest of the proof?
+
+Current repo answer:
+- yes on explicitization
+- no on comfort: once the bound is written out, the large-prime tail is still too expensive at
+  the public candidate thresholds unless the tail treatment or the truncation parameter improves
+- see `LEMMA21_EXPLICIT_BOUND.md`
+
+### B. Lemma 2.2 progression error
+
+Need:
+- explicit control of the tail over primes dividing `ab+1`
+- a quantitative replacement for the `<< N / sqrt(log N)` term
+
+Route question:
+- is this the actual bottleneck, as the public forum discussion suggests?
+
+### C. Choice of `eta`
+
+Need:
+- identify where the proof requires a small absolute `eta`
+- determine whether the proof ever needs a specific lower ceiling on `eta`, or only that `eta` be less than the smallest surviving numerical slack after error absorption
+
+### D. Final case assembly
+
+Need:
+- for each case, record the exact main-term bound and the admissible total error budget
+- verify which branch determines the final threshold
+
+Current belief:
+- the `0.0377` case is the active bottleneck until a sharper calculation says otherwise
+
+## Work order
+
+1. Freeze the exact or conservatively certified main-term slack for the `0.0377` branch.
+2. Extract every hidden constant and error input from Lemma 2.1.
+3. Decide whether the resulting large-prime tail can fit inside the weakest-branch slack at any plausible threshold scale.
+4. If not at `T = floor(sqrt(log N))`, scan larger truncation parameters before demanding a deeper new theorem.
+5. Extract every hidden constant and error input from Lemma 2.2 using the best currently justified truncation choice.
+6. Only then decide whether to keep pushing analytic extraction or pivot harder into bounded finite computation.
+
+Current repo posture:
+- Step 1 is now complete at a conservative numerical level.
+- Step 2 is now complete at a one-sided explicit level.
+- Step 3 is now complete: the route should enlarge `T` before asking for a deeper tail theorem.
+- Step 5 is now complete at a witness-budget level.
+- The next unresolved work is no longer mathematical: it is deciding how to commit and review
+  the surfaced candidate package.
+
+## What would count as progress
+
+- a line-by-line list of every non-explicit estimate in Lemma 2.1 and Lemma 2.2
+- an exact or conservative usable-slack budget for the weakest branch
+- an explicit declaration of which branch sets the threshold
+- a clean budget statement for how much of the roughly `0.002388` branch slack is available to Lemma 2.1, Lemma 2.2, and `eta`
+- a clear statement of whether the large-prime tail or the small-prime discretization is the
+  real bottleneck
+- a clear statement of whether the proof should first change `T` or first seek a sharper tail
+  theorem
+- a line-by-line weakest-branch assembly for at least one concrete witness value of `T`
+- a line-by-line ledger for the surviving `eta` room after all lemma-level costs
+- a branch comparison ledger showing whether the weakest-branch witness already dominates the
+  rest of the proof
+- a proposition-level explicit witness note saying exactly what the repo now claims and what it
+  still does not claim
+- a proof-obligation ledger saying what still has to be frozen before a public-truth update
+- a certified numerical ledger for every decimal input used in the current witness
+- a theorem-style note that assembles the current witness in one proof-shaped artifact
+- a publication handoff decision for whether that theorem-style note belongs in the paper
+  bundle, a public review artifact, or both
+- a short share-ready summary for maintainers or public-facing notes
+- a justified statement like:
+  - “the current proof architecture plausibly yields an explicit threshold”
+  - or
+  - “the current proof architecture is too lossy without a new lemma or better error term”

package/packs/number-theory/problems/848/FRONTIER_NOTE.md

@@ -0,0 +1,48 @@
+# Problem 848 Frontier Note
+
+Problem 848 is a decidable finite-check problem, not a fresh asymptotic frontier.
+The optimization target is full finite-gap closure, not "smallest `N0`" in isolation.
+
+The live route is `finite_check_gap_closure`:
+- keep Sawhney's sufficiently-large-`N` theorem exact
+- do not widen `decidable` into `solved`
+- treat explicit-threshold extraction and finite verification as separate but coupled lanes
+
+Imported public threshold timeline:
+- `N0 = 7 x 10^17` on 2026-03-21
+- `N0 = 3.3 x 10^17` on 2026-03-22
+- `N0 = 2.64 x 10^17` on 2026-03-23
+
+So the repo's current `exp(1420)` candidate should be read as an audited workspace artifact,
+not as the best public threshold.
+
+The smallest honest next move is not “solve 848”.
+It is:
+- keep the current repo candidate claim-safe while turning it into a review unit
+- maintain the distinction between:
+  - theorem-style repo candidate
+  - public review artifact
+  - finished publication proof
+- use the current explicit witness package as the handoff surface, not as a reason to
+  silently upgrade the problem to `solved`
+
+The current package is already internally reviewed enough for handoff:
+- the paper bundle now has drafted introduction, preliminaries, and related-work sections
+- the surfaced candidate package has no remaining placeholder text
+- tests and publish-surface checks are green
+
+Chosen next lane:
+- bounded finite verification under the best imported threshold currently tracked
+
+Why this lane wins the next cycle:
+- the real objective is to close the finite remainder, not just publish a smaller `N0`
+- imported threshold progress already exists, so the repo needs an interval-certification
+  surface to make use of it
+- the public thread already contains one verification attempt that was later corrected and
+  criticized as difficult to verify, so trust and reproducibility have to be first-class
+
+Read next:
+- `BOUNDED_VERIFICATION_PLAN.md`
+- `VERIFICATION_REGIMES.md`
+- `VERIFICATION_CERTIFICATE_SPEC.md`
+- `EXTERNAL_VERIFICATION_LEDGER.md`

package/packs/number-theory/problems/848/LEMMA21_EXPLICIT_BOUND.md

@@ -0,0 +1,200 @@
+# Problem 848 Lemma 2.1 Explicit Bound
+
+This note closes the current `N848.G1.A4` task:
+- replace the hidden `<<` and `N^(o(1))` steps in Lemma 2.1 with an explicit one-sided bound
+- identify which remainder term is actually binding for the `0.0377` branch
+
+Scope:
+- this is still not a full explicit threshold proof
+- this note only packages the first explicit Lemma 2.1 remainder bound
+- the goal is to decide what the next honest analytic move must be
+
+## Source surface
+
+Primary source:
+- Sawhney, `Problem_848.pdf`, Lemma 2.1, pages 1-2
+
+Support source:
+- `erdosproblems.com/forum/thread/848`
+
+## Setup
+
+Let
+- `U(P; N, q, t)` denote the set of `n in [N]` such that `n ≡ t mod q` and
+  `n mod p^2 in R_p` for at least one `p in P`
+- `|R_p| <= 2`
+- `R_p = empty` whenever `(p, q) != 1`
+- `T = floor(sqrt(log N))`
+
+Write:
+- `P_<= = {p in P : p <= T}`
+- `P_> = {p in P : T < p <= N^(1/2)}`
+- `x_p = |R_p| / p^2`
+- `m = |P_<=|`
+
+The lemma in the note states
+- `|U(P; N, q, t) - (N/q) * (1 - prod_{p in P} (1 - x_p))| << N / sqrt(log N)`
+
+For the route, the important question is not the asymptotic statement itself. It is:
+- what explicit upper bound can replace the hidden remainder when we only need a one-sided
+  inequality for the density casework?
+
+## Explicit large-prime tail
+
+For each `p in P_>` and each `r in R_p`, the congruence conditions
+- `n ≡ t mod q`
+- `n ≡ r mod p^2`
+
+define at most one residue class modulo `q p^2`, because `(p, q) = 1` whenever `R_p` is
+nonempty. Therefore
+
+`|{n in [N] : n ≡ t mod q and n ≡ r mod p^2}| <= N / (q p^2) + 1`.
+
+Summing over the at most two allowed residues gives
+
+`|{n in [N] : n ≡ t mod q and n mod p^2 in R_p}| <= 2N / (q p^2) + 2`.
+
+By a union bound over `p in P_>`,
+
+`|U(P_>; N, q, t)| <= (2N/q) * sum_{p in P_>} 1/p^2 + 2|P_>|`.
+
+So the proof's hidden large-prime tail is explicitly bounded by
+
+`(2N/q) * sum_{p in P, p > T} 1/p^2 + 2 pi(N^(1/2))`.
+
+This is the first source-backed place where the route can choose between:
+- the original crude integer tail `sum_{n > T} 1/n^2 < 1/(T - 1)`
+- a sharper prime-only tail, which the public forum discussion identifies as the more
+  realistic path
+
+## Explicit small-prime inclusion-exclusion remainder
+
+Consider the truncated union over `P_<=`.
+
+For each nonempty subset `S subseteq P_<=`, inclusion-exclusion introduces the intersection
+
+`I_S = | intersection_{p in S} {n in [N] : n mod p^2 in R_p} intersection {n in [N] : n ≡ t mod q} |`.
+
+For a fixed choice of one residue from each `R_p`, the congruence system determines exactly
+one residue class modulo `q * prod_{p in S} p^2`. Hence
+
+`I_S = (N/q) * prod_{p in S} x_p + E_S`
+
+with
+
+`|E_S| <= prod_{p in S} |R_p| <= 2^{|S|}`.
+
+Summing these errors through inclusion-exclusion gives
+
+`|U(P_<=; N, q, t) - (N/q) * (1 - prod_{p in P_<=} (1 - x_p))| <= sum_{empty != S subseteq P_<=} 2^{|S|}`
+
+and therefore
+
+`|U(P_<=; N, q, t) - (N/q) * (1 - prod_{p in P_<=} (1 - x_p))| <= 3^m - 1`.
+
+This is the explicit replacement for the note's `N^(o(1))` step.
+
+## One-sided explicit Lemma 2.1 bound
+
+Combining the truncated inclusion-exclusion bound with the large-prime union bound yields
+
+`|U(P; N, q, t)| <= (N/q) * (1 - prod_{p in P} (1 - x_p)) + (3^m - 1) + (2N/q) * sum_{p in P, p > T} 1/p^2 + 2 pi(N^(1/2))`.
+
+This is weaker than the asymptotic note in two ways:
+- it is only one-sided
+- it keeps discrete counting terms instead of burying them in `o(1)`
+
+But for the route, it is enough:
+- the casework only needs an upper bound
+- the new expression cleanly separates the tail term from the small-prime discretization
+
+## Specialization to the `A*` branch
+
+In the weakest `0.0377` branch of the proof:
+- `q = 25`
+- there are `23` residue classes modulo `25` contributing to `A*`
+- `P = {p prime : p ≡ 1 mod 4 and p >= 13}`
+- each active prime contributes exactly `|R_p| = 2`
+
+So the explicit density bound becomes
+
+`|A*| / N <= (23/25) * (1 - prod_{p ≡ 1 mod 4, p >= 13} (1 - 2/p^2))`
+
+plus the explicit remainder
+
+`(23/N) * (3^m - 1) + (46/25) * sum_{p > T, p ≡ 1 mod 4} 1/p^2 + 46 * pi_{1 mod 4}(N^(1/2)) / N`
+
+where
+- `m = |{p <= T : p ≡ 1 mod 4 and p >= 13}|`
+
+## What the explicit bound says at public threshold candidates
+
+The small-prime remainder is not the live problem.
+
+For the public candidate scales mentioned on the forum:
+
+### 1. `N = exp(1420)` so `T = floor(sqrt(log N)) = 37`
+
+Numerical support:
+- `m = 4`, so the discrete inclusion-exclusion term is exactly `23 * (3^4 - 1) / N = 1840 / N`
+- summing `1/p^2` over primes `p ≡ 1 mod 4`, `p > 37`, up to `5,000,000`, then adding the
+  crude tail envelope `1 / 5,000,000`, gives about `0.00251707`
+- multiplying by the branch coefficient `46/25` gives a tail-density penalty about
+  `0.00463141`
+
+### 2. `N = exp(1958)` so `T = floor(sqrt(log N)) = 44`
+
+Numerical support:
+- `m = 5`, so the discrete inclusion-exclusion term is exactly `23 * (3^5 - 1) / N = 5566 / N`
+- the same prime-tail computation gives about `0.00192219`
+- multiplying by `46/25` gives a tail-density penalty about `0.00353682`
+
+### 3. Comparison with the frozen weakest-branch slack
+
+From `WEAKEST_CASE_BUDGET.md`, the current branch-level slack before analytic error
+absorption is about `0.00238869`.
+
+Therefore:
+- the small-prime discretization term is negligible at these scales
+- the large-prime tail already exceeds the full branch slack at both `exp(1420)` and
+  `exp(1958)`
+- this happens before paying for Lemma 2.2 or for the final `eta` absorption
+
+If one keeps the proof's even cruder integer-tail estimate
+- `sum_{n > T} 1/n^2 < 1/(T - 1)`
+
+then the `A*` density penalty alone is
+- about `0.05111` at `T = 37`
+- about `0.04279` at `T = 44`
+
+which is completely incompatible with the frozen branch slack.
+
+## Honest route consequence
+
+This closes the current question behind `N848.G1.A4`.
+
+The new exact route reading is:
+- Lemma 2.1 is now explicit enough to see the structure of the loss
+- the `N^(o(1))` / inclusion-exclusion discretization is not the real blocker
+- the large-prime tail is the live analytic bottleneck
+- keeping `T = floor(sqrt(log N))` and the crude tail treatment is too lossy for the
+  public threshold candidates discussed so far
+
+So the next honest move is no longer
+- “make Lemma 2.1 explicit”
+
+It is:
+- improve the large-prime tail or the truncation parameter before spending time on more
+  bookkeeping
+
+## Numerical method note
+
+The prime-tail support numbers above are not yet formal proof artifacts.
+
+They were obtained by:
+- summing over all primes up to `5,000,000`
+- restricting to the `1 mod 4` primes for the `A*` branch
+- adding the crude tail envelope `sum_{n > 5,000,000} 1/n^2 < 1 / 5,000,000`
+
+This is conservative enough for route guidance, but not yet the final explicit-threshold
+certificate.