Replay the complete PR3605 wrapper with a fresh retained bootstrap

[roger-sato]

Summary

Complete the previously unexecuted public full-wrapper path from PR3605 at Ross Richardson's exact head 3efbb5de06a49ffeea6684482f2d3fe48cdca3fb, answering its open replay handoff.

• Run the original wrapper with --jobs 1, no --retained and no --prepare-only. Authenticate all 157 supplied source/input files before and after the run. No predecessor source or checker was patched.
• Retain a compact transcript and bounded normal/-O verifier covering the ordered stage inventories, raw logs/scripts, complete source/cache inventories, four bootstrap native chains, fresh 16-stage parent admission, two changed native chains, final release chain and exact existing endpoint/profile.
• Explain the different generated manifest: 648 file bytes match the original public inventory, and exactly four JSON hashes are reproduced by restoring two phase receipt hashes and their PLAN/SPEC references. The local collector checks actual large PLAN bytes; the public bounded checker restores the three smaller retained JSONs and checks the recorded PLAN comparison, with that limitation explicit.
• Distinguish no external retained-cache reuse from the expected reuse of caches generated earlier in this same run. Retained ancestral geometry and all-size theorem hypotheses remain. The bounded verifier checks recorded evidence; it does not rerun the solvers or reauthenticate omitted large artifacts.

The published endpoint is unchanged: 0.001730006549538841847005 - 10^-46 - 10^-70, with 2,717,280 roles, 6,660 calls and 740 replicas. This is finite reproduction evidence, not a new κ, unconditional/formal-kernel theorem, or selected-frontier change. The original controller summary of 15 is preserved; actual ASYNC-AUDIT and generated SPEC require 25 controls.

The task scope explicitly changes from the portable-replay card to this live unfilled validation handoff. PR3659 is an adjacent histogram/root comparison; this PR supplies the full source/native/owner/physical/prefix/outer execution path. Original author and inherited source credits are preserved, with substantial OpenAI Codex assistance disclosed.

Test plan

• Exact original full wrapper exited 0 with PASS_STAGED_PAIR_COMPLETE_CONDITIONAL_REPLAY in 2,443.539 seconds (40m 43.5s), from 2026-10-18 13:20:59 UTC to 14:01:42 UTC. All 18 top-level stages passed; sources were unchanged and stderr was empty. Final release receipt: efc62e10c52ffd5170712d9f9edc00d79bedf73830a5111dbdfd7493726bcdc6; generated manifest: 7a94445290bc20adecb1c8a378c73987a71aade611c1ef5b7e295ff037e674b6.
• Local collector authenticated all source, generated-package and native-cache bytes, native word payload identities, and all four exact historical JSON reconstructions. The compact transcript has 214 members / 880,507 bytes; archive SHA-256 eb7b68908640192e6923a8bd21616bc70372ebbe7f1bb06a48b53f37817eafa6.
• Python 3.12.14 scripts/research_package.py --package research/staged-pair-fresh-bootstrap-replay --check-only and the registered normal/optimized bounded replays passed, including all 15 semantic rejection controls. These checks do not rerun the 40-minute computation.
• Separate AI review checked the actual source/bootstrap/native/parent/terminal receipts, raw owner/arithmetic leaves, three exact rational moment-engine enclosures, exact endpoint/root/profile, and all 157 original supplied inputs. This is independent execution review, not independent human or formal-kernel review.
• Separate fixture review passed normal/optimized execution with warnings treated as errors, all 15 built-in controls, and 14 additional actual CLI checks (positive plus six envelope-rebound mutations in both modes).
• Package metadata/source pins and publication content scan passed; no credential-pattern matches across all 214 retained members.
• The publication diff is limited to nine new package files. check_research_commit.py --ref HEAD --check-source-history passed on publication commit 9f7813943b488d386407b1f4fdc1a3b2a8d39f84 in a separate full-history checkout; all nine committed file bytes and declared source ancestry matched.

Fork-hosted bounded CI — 2026-10-18

The Registered research packages run passed on Ubuntu 24.04.5 x64 / CPython 3.12.15 at this exact publication head, 9f7813943b488d386407b1f4fdc1a3b2a8d39f84, using the unchanged workflow and base 20ecf2891d80b5acf18d980b2164fb44ea7faa90. No source or workflow change was needed.

• The plan job passed all 254 executed onboarding helper tests, with 11 Windows-only skips, and selected exactly research/staged-pair-fresh-bootstrap-replay.
• Both registered commands (normal Python and -O) passed, each including all 15 semantic rejection controls. The runner reported PASS_REGISTERED_RESEARCH_PACKAGE; the subsequent git diff --exit-code check also passed.
• This establishes Linux portability of the bounded retained-evidence checker. The 40-minute scientific computation remains the separately recorded macOS run above.

The upstream PR run at the same head still reports action_required and needs maintainer approval. The fork run provides additional execution evidence; it does not replace that upstream approval/check.

Reproduce the bounded hosted check:

gh workflow run research-packages.yml --repo roger-sato/integer-mult-bounds \
  --ref audit/staged-pair-fresh-bootstrap \
  -f base=20ecf2891d80b5acf18d980b2164fb44ea7faa90