[Clear1nsight]

Change and mathematical scope

Add research/h25-v03-prefix-stage-preparation: the byte-unchanged frozen prepare_prefix_stage(request) source, exact failed/intermediate/final history, and a registered replay of its 22 NEW preparer methods per normal/optimized profile. It actually owns descriptor-relative installation, installs P -> Q -> M -> JOBS/STAGE -> configuration/policy/selection DATA, and records its final attempted-operation bill after consumed cleanup. It never launches its output. The whole-DAG tests use explicitly inert fixture pins with the frozen restoration contexts.

The connected prerequisite source components are accepted within their conditional contracts. No genuine complete compiler-input closure, full bootstrap-control campaign, C07 interval, source/native/runtime/funding admission, complete physical fit, caller/parent exit, paid-profile gain or kappa improvement is established. Outside costs remain OPEN_NOT_ZERO. The app goal is separately paused; guarded 94.550533479/2400 is unapproved and is not a physical total.

Task and overlap

Original source investigation: PR #6321. Frozen H25 benchmark: PR #6029. Immediate predecessor: PR #6371. Companion caller: PR #6373.

Canonical preparation base 7c49fe32c55de02c2c873655c804b16e4c3fef0a. This explicit source-publication assignment has changed scope from the provenance-only portable-replay catalogue card; no new scientific task was drawn.

Nearest supplied checkpoints: PR6371, entry head 8e772eadcbdee11289eb2eac887ffc15b864587c, and PR6370, inputs head e144f3b0a440fd33efdcfa4e0bd1af059dfd4fb3. This offline preparation does not refresh their hosted state. The separate bounded caller PR6373, exact head b91c5c74f252c1a98ab8344d012702170004ed13, is submitted; the exact 34,716-byte caller SHA256 25138c2c998762649c6a61cb8a6ae8aacd81bd533591b082ab6f58def712fca2 is retained and hashed as inert archive data. Its separate 11 methods are excluded from this replay and count. No predecessor package is duplicated for provenance.

The export preserves source snapshots 75bc, f8a, a2b and final 25c, pre-freeze self-correction, initial failed counterexamples, whole-DAG oracle count correction from 21 to 20 and failing v1 outputs, exact final selectors, author receipt, final P2/P4 reviews, designs/inputgate/repairwindow/sourceclosure/freezetiming, operation bounds, and raw outputs from the source round. Earlier 21-method logs are not relabelled as final 22-method evidence. Exact original entry/Group01 codec intersection is preserved; this does not prove every original 8,191-row Q fits.

Validation

Python 3.12, standard library only, writable POSIX/Linux temporary filesystem. Observed local commands:

python3 -B scripts/research_package.py --package research/h25-v03-prefix-stage-preparation --check-only
python3 -B scripts/research_package.py --package research/h25-v03-prefix-stage-preparation

Metadata/pins pass. Registered normal and optimized runs each passed all 22 methods pass, each fixture in a fresh isolated interpreter. A package-only copy, with no Git or original checkout dependencies, passes both direct commands:

python3 -B verify.py --profile normal
python3 -O -B verify.py --profile optimized

99 workspace-source exports and the two base license/notice files preserve exact bytes; source/test AST identity is checked where applicable. All 108 package files are explicitly inventoried. Added/unknown-host manifest parts fail before any fixture runs. A first publication-wrapper path-prefix failure was preserved externally and repaired only in the wrapper binding; the frozen core and fixtures were untouched. Validation outputs are outside the package. The original/native/full entry/full caller/compiler/namespace/Popen/domain/target suites were not executed.

Actual no-index whitespace diagnostics retain EOF blank lines in four frozen archived Markdown files; these bytes were deliberately preserved. The isolated local cache lacks an unrelated PDF blob and is not represented as complete original source history. The reviewed bytes are committed with tree 93be678d9b422e5d94343b7eaff26a226861ffe8, sole base parent 7c49fe32c55de02c2c873655c804b16e4c3fef0a and remote publication head b02bfe3338f7fe57818262b33f8e5c77ac0f1c93. Required committed-byte/source-history/executable-mode checks passed for all 108 files; history coverage is the declared base/card/rules, not complete original scientific history. Remote and local commits have the identical tree; publication metadata/message/timestamp differ. Hosted exact-head checks are pending and will be read back before delivery is complete.

Credit

Preserve Chris R, Clear1nsight, all original authors and source notices, PR6321/6029/5864, unpaidmoon/Cursor/Grok PR6339, and Douglas Colkitt. Exact base LICENSE/NOTICE are included. Substantial OpenAI Pro/Codex/GPT-6.1 Sol Max assistance and agent reviews are disclosed; these are not human or proof-kernel acceptance.

Open next questions

See the package Open next questions and structured TASK.json: obtain independently admitted genuine complete compiler-input closure/all-nine-category rows and exact source/stage/command/copy/funding/execution records, then compare to the frozen request schema. Separately establish original T0/native/caller premises, C07 interval/ancestor stability and complete physical CPU/wall/memory/I/O/storage/approval/final-tail costs/funding. These require missing external evidence and separate execution authority. The source editing window is closed and this export excludes all new composition debug work.

[Clear1nsight]

PR6374 head b02bfe3338f7fe57818262b33f8e5c77ac0f1c93; 108 files; 22 NEW preparer fixture methods per profile. Hosted run 41469024302.

Exact-head source delivery is verified: both actual planning and explicit package replay, including unchanged committed inputs, passed on attempt 1. All exported blob identities and the remote tree match the frozen independently reviewed package. The PR remains a draft, with maintainer edits off. The title formatting artifact was corrected without changing the head or package bytes.

Canonical technical handoff and existing coordination-hub reply have been posted and read back. They retain the exact sources, failures, reviews, reproduction commands, limits and attribution.

This closes this conditional source/export contribution only. Full original compiler-input closure and bootstrap-control campaign require the genuine approved determinant/channel inventory and immutable lifetime, current source/native/stage/command/funding/execution inputs, intact original baseline and named corruption/current-custody controls. Those original operations remain unexecuted. App goal separately observed paused; guarded attempt unapproved; recorded policy debit is not a complete physical invoice, and outside/final costs remain OPEN_NOT_ZERO. Preserve original authors/licenses and substantial Pro/Codex/GPT-6.1 Sol Max assistance.

Original H25 benchmark: PR #6029. Original source investigation: PR #6321.