Bound staged-supplier packing for every replica count

[RossRichardson]

Summary

An independently certified 2,551-pattern dual proves
C < .001730096363528623534231 for every positive integer replica count with
the staged 644/660 supplier's fixed capacities, helper widths, local bill,
dimension 73 and live stock 896T. A complete released five-owner T740 witness
supports κ=.001730006690311348872326−10^-46−10^-70, a small packing gain over
PR3605, with all owner, global invoice and conditional-transfer costs retained.

Final overlap found the stronger concurrent PR3624
at 515a7c4490aa32ac3ca5014feb6f7da0d6952863, reporting
κ=.001730047552093344159584−10^-46−10^-70 at T450. Its exact certificate and
invoice agree with this ceiling's fixed-inventory premises. We inspected those
premises but did not rerun its full admission. This PR's primary contribution is
the independent all-T ceiling; the five-owner witness is not current best.

Scope and limits

All 2,717,280 owner rows are regenerated and independently audited; local words,
source order, dispatch, 408 helper roles and 91 private buffers are unchanged.
A new 2,551-pattern dual bounds this fixed-inventory class below
.001730096363528623534231 for every positive integer T with the stated fixed
data. Chronology/family attainment of that relaxation is unknown; no global
packing optimum or cap on changed constructions is claimed.

The source closure retains 36 exact original PR3605 blobs with SHA256 and original
Git blob identities. Generated gzip representations are admitted by complete
payload semantics. The bounded check is separate from the provided long cold
recipe and completed actual admission. Both unchanged current native providers
and retained parent evidence are authenticated reuse; fresh global release
checks include all 8,758 actual charts and 25 controller controls. Actual prefix
is 1,479 bits, sufficient reserve 1,512, paid budget 2,048; physical stock 663,549.
The initial pre-check output-directory collision and successful unchanged retry
are preserved explicitly. All inherited all-size, uniform cost,
physical hosting, mutable-source, rounding, tape, bootstrap, PR559 recovery,
analytic transfer and joint BIT hypotheses remain conditional. No selected
record change or formal proof is claimed.

Test plan

• Metadata and exact live source pins.
• Optimized-parent registered verifier: fresh complete literal ownership,
  full histogram, three rigorous moments, 63 modern/17 rejecting and 47 bootstrap/4
  controls; 85.5 seconds locally.
• All 2,551 fixed-inventory patterns, with 2,539 strict reduced costs and 12
  symbolic zero costs, plus adjacent relaxed-root signs.
• Complete original strict release (1,352.8 seconds) and independent final
  source/literal/invoice/outer binding; exact rational improvement over PR3605.
• Optimized evidence-binding check against the fresh literal and arithmetic.
• Exact committed-byte and strict source-history preflight: 126 files.
• Committed Git-free isolated/optimized replay: 78.1 seconds; all 126 source
  and package files unchanged, including fresh literal/dual and release binding.
• Automatic hosted CI passed at exact head bcac7fe04b9b7f7f5d3c27d620561f96c4ae5608, including the bounded replay and unchanged-input check.

Sources and credit

Ross Richardson with substantial OpenAI Codex assistance and separate AI reviews.
Exact predecessor: PR3605
at 3efbb5de06a49ffeea6684482f2d3fe48cdca3fb. Credit remains with dcharb78's
PR3133 carrier folds, Jeremy Diviney's PR3148 owner/cap work, the PR3531
positive-binomial engine, and all named supplier/admission/owner predecessors
whose original notices are retained. The standalone all-row audit and dual are
adapted from separate reviewers, not the producer. Coordination remains in
Discussion3606.

[RossRichardson]

Scoped hygiene follow-up at bcac7fe04b9b7f7f5d3c27d620561f96c4ae5608: independent_owners.py inherits the five ineffective rejection controls identified in the #3605 review. They test local predicates without sending malformed input through the all-row validator; disabling the relevant gates leaves these controls green.

A narrow local patch factors the existing checks and tests actual journal/profile mutations with exact rejection reasons. It has not been pushed; this issue remains open. The real all-row gates, including the active-local-pair exclusion, remain present, and this finding is not an arithmetic counterexample.

The PR's bounded scope remains appropriate: the all-replica ceiling applies to the explicitly fixed inventory, and the five-owner witness is already described separately from the stronger concurrent result. This control repair should stay within that scope and preserve the distinction between historical release evidence and validation at a repaired head.