feat: publish the complete exact-side catalogue on the web - #435

Review A (senior review, round 1) · reviewed head
760dd9538a4b2c7c957ce6950d4624d790da3400 · base
6a9b2876d40ebbc4d709ecfecd1bf107e6db95cb.

Reviewer: strong tier, requested gpt-6-astra at xhigh. Trusted channel: independent
review prepared for formal GitHub COMMENT; author jlevy is OWNER in the same
repository. Publication remains coordinator-owned.

Current immediate-layer inventory: 144 changed paths.
Coverage is composed from the explicit prior full audit and the reviewed successor delta
below; unchanged source, byte-only checks and skipped execution are distinguished.

Verdict: changes required for current own-layer C1 and the open foundation findings.
The earlier diagnostic A1 is fixed: README/SYNOPSIS now use 320 exact / 4 numeric /
170 historical and the source packet’s 54-row census is explicitly dated.
The earlier diagnostic child D1 is fixed: the hosted-PDF migration now preserves
source-proof annotations through manifest-declared original archive identities.
Do not publish those resolved diagnostic findings as open current findings.
Parent #403 A1/B1/D1/D2 still need resolution and qualified landing.

This is the first formal four-role review prepared for this PR. The earlier 34f7
diagnostic drafts were not published.
It bridges the complete prior 142-path audit (133 exact path bridges plus nine direct
changed-path reads) and the independently retained full source audit.
The preceding 34f7 to 1c102f4f 54-path successor delta has 18 exact main91ca byte
bridges and 36 own changed paths, including parent changes; complete own changes,
callers and records were assessed.
The complete inventory distinguishes direct reads, prior semantic coverage and byte
bridges rather than claiming all unchanged bytes were reread.

The final three-path delta from 1c102f4f to this head was read in full: Pages now
retains the exact archived KKT-results leaf needed by the renderer, the real Git
sparse-checkout contract checks that leaf and neighboring exclusions, and the work plan
distinguishes two reported input downloads from unexecuted solver/search work.
All source data and generated register bytes match 1c102f4f. No new tests or external
input retrieval were executed for this final bridge.

All 170 historical register entries and 291 current entries are identical to the
reviewed earlier register.
The remaining 33 current rows have the same custody-checked finite-rational projection
reviewed and tested in the parent: status, side, lower bound and KKT fields remain
unchanged.
Current totals are 320 exact identities, four numeric rows (29,55,71,105), 170
historical entries and 77 proved cases.
The degree 672 source polynomial, its 673 coefficient strings, all 38 SVG joins, six
article identities and 201 printed occurrences forming 182 coefficient/side pairs retain
their original independent source-audit scope.
The source-stated root ordinal is not independently proved.

The 13 original PDFs (198,944,687 bytes) retain the prior live release 407039473 /
data/source-pdfs-v1 asset identity checks and manifest linkage; they were not downloaded
again. Archive raw bytes, licenses and source credit remain preserved.
No author program, Sage/native build, new geometry replay or full child gate was
executed in this refresh.
The parent register’s 53 independent tests passed; child semantic comparison covers all
324 rows, and all 170 historical entries match.
This is composed source review, not new full mathematical replay.

Fresh GitHub discovery confirms OWNER jlevy, current child head and parent base, zero
formal reviews, zero inline comments and zero PR comments.
Earlier complete related-issue discovery remains in the audit receipt.
Publisher must repeat that guard.
At this head, Packing required and the exact-side-values job pass; Pages publish and
Pages required fail, while the full run is still active at this review’s observation.
These are distinct outcomes; old failing or passing runs cannot be relabeled.
The foundation must be fixed and requalified before this layer can land.

Review B (security review, round 1) · reviewed head
760dd9538a4b2c7c957ce6950d4624d790da3400 · base
6a9b2876d40ebbc4d709ecfecd1bf107e6db95cb.

Reviewer: strong tier, requested gpt-6-astra at xhigh. Trusted channel: independent
review prepared for formal GitHub COMMENT; author jlevy is OWNER in the same
repository. Publication remains coordinator-owned.

Current immediate-layer inventory: 144 changed paths.
Coverage is composed from the explicit prior full audit and the reviewed successor delta
below; unchanged source, byte-only checks and skipped execution are distinguished.

Verdict: changes required through inherited parent B1 (High). The child’s
linear/precision fixes do not gate run_svg --write-record on a matching source
acquisition digest, so the same source-custody admission defect remains.
Keep its identity as parent B1 rather than inventing a duplicate child finding.
The hosted-paper parser maps only manifest-declared release URLs to original
packing/resources/papers paths; the positive local/hosted cases and unrelated-URL
refusal preserve source identity.
No license, raw-source or source-credit relaxation is needed.
Current publication C1 also requires a reviewed script classification while preserving
unknown-script rejection; it is recorded once in Review C.

This is the first formal four-role review prepared for this PR. The earlier 34f7
diagnostic drafts were not published.
It bridges the complete prior 142-path audit (133 exact path bridges plus nine direct
changed-path reads) and the independently retained full source audit.
The preceding 34f7 to 1c102f4f 54-path successor delta has 18 exact main91ca byte
bridges and 36 own changed paths, including parent changes; complete own changes,
callers and records were assessed.
The complete inventory distinguishes direct reads, prior semantic coverage and byte
bridges rather than claiming all unchanged bytes were reread.

The final three-path delta from 1c102f4f to this head was read in full: Pages now
retains the exact archived KKT-results leaf needed by the renderer, the real Git
sparse-checkout contract checks that leaf and neighboring exclusions, and the work plan
distinguishes two reported input downloads from unexecuted solver/search work.
All source data and generated register bytes match 1c102f4f. No new tests or external
input retrieval were executed for this final bridge.

All 170 historical register entries and 291 current entries are identical to the
reviewed earlier register.
The remaining 33 current rows have the same custody-checked finite-rational projection
reviewed and tested in the parent: status, side, lower bound and KKT fields remain
unchanged.
Current totals are 320 exact identities, four numeric rows (29,55,71,105), 170
historical entries and 77 proved cases.
The degree 672 source polynomial, its 673 coefficient strings, all 38 SVG joins, six
article identities and 201 printed occurrences forming 182 coefficient/side pairs retain
their original independent source-audit scope.
The source-stated root ordinal is not independently proved.

The 13 original PDFs (198,944,687 bytes) retain the prior live release 407039473 /
data/source-pdfs-v1 asset identity checks and manifest linkage; they were not downloaded
again. Archive raw bytes, licenses and source credit remain preserved.
No author program, Sage/native build, new geometry replay or full child gate was
executed in this refresh.
The parent register’s 53 independent tests passed; child semantic comparison covers all
324 rows, and all 170 historical entries match.
This is composed source review, not new full mathematical replay.

Fresh GitHub discovery confirms OWNER jlevy, current child head and parent base, zero
formal reviews, zero inline comments and zero PR comments.
Earlier complete related-issue discovery remains in the audit receipt.
Publisher must repeat that guard.
At this head, Packing required and the exact-side-values job pass; Pages publish and
Pages required fail, while the full run is still active at this review’s observation.
These are distinct outcomes; old failing or passing runs cannot be relabeled.
The foundation must be fixed and requalified before this layer can land.

Review C (performance review, round 1) · reviewed head
760dd9538a4b2c7c957ce6950d4624d790da3400 · base
6a9b2876d40ebbc4d709ecfecd1bf107e6db95cb.

Reviewer: strong tier, requested gpt-6-astra at xhigh. Trusted channel: independent
review prepared for formal GitHub COMMENT; author jlevy is OWNER in the same
repository. Publication remains coordinator-owned.

Current immediate-layer inventory: 144 changed paths.
Coverage is composed from the explicit prior full audit and the reviewed successor delta
below; unchanged source, byte-only checks and skipped execution are distinguished.

Verdict: changes required for C1; current qualification and input scope remain explicit.

C1 (High), publication contract incomplete. At this exact head, Pages publish
job 113591079326
passes 5,321 of 6,332 checks.
Its 1,011 failures comprise 1,005 unregistered outputs (508 coefficient JSON, 494
metadata JSON, the index JSON, browser JS and complete HTML), the 5,016,486-byte
complete HTML exceeding the 2,000,000-byte hard limit, four inline startup scripts
exceeding 4,096 bytes, and the unclassified browser script.
site_urls.declared_outputs and the retained URL register must declare all actual
producer outputs; renderer packaging must satisfy the existing HTML and startup limits,
and the browser program needs a reviewed declaration matching its behavior.
Keep unknown-output/script rejection, complete scientific content and the existing
limits intact. Validate the assembled publication, including negative controls; the
individual exact-side-values job already passes.

The current selector counts 201,286,474 bytes, 40,118 below the 192 MiB cap; the
unchanged main selector would count 204,957,003. Both stack layers omit the same 312
historical paths/3,670,529 bytes.
This is a current numeric pass with narrow headroom, not proof of all transitive input
completeness. Known consumers and protected replay inputs were traced without a concrete
lost deciding input; lossless generated gzip preserves the complete older input set and
offers more margin. The PDF removals contribute zero snapshot bytes because those files
were already excluded.
Earlier browser payload reductions are raw-byte measurements, not new latency results or
current hosted gates.

This is the first formal four-role review prepared for this PR. The earlier 34f7
diagnostic drafts were not published.
It bridges the complete prior 142-path audit (133 exact path bridges plus nine direct
changed-path reads) and the independently retained full source audit.
The preceding 34f7 to 1c102f4f 54-path successor delta has 18 exact main91ca byte
bridges and 36 own changed paths, including parent changes; complete own changes,
callers and records were assessed.
The complete inventory distinguishes direct reads, prior semantic coverage and byte
bridges rather than claiming all unchanged bytes were reread.

The final three-path delta from 1c102f4f to this head was read in full: Pages now
retains the exact archived KKT-results leaf needed by the renderer, the real Git
sparse-checkout contract checks that leaf and neighboring exclusions, and the work plan
distinguishes two reported input downloads from unexecuted solver/search work.
All source data and generated register bytes match 1c102f4f. No new tests or external
input retrieval were executed for this final bridge.

All 170 historical register entries and 291 current entries are identical to the
reviewed earlier register.
The remaining 33 current rows have the same custody-checked finite-rational projection
reviewed and tested in the parent: status, side, lower bound and KKT fields remain
unchanged.
Current totals are 320 exact identities, four numeric rows (29,55,71,105), 170
historical entries and 77 proved cases.
The degree 672 source polynomial, its 673 coefficient strings, all 38 SVG joins, six
article identities and 201 printed occurrences forming 182 coefficient/side pairs retain
their original independent source-audit scope.
The source-stated root ordinal is not independently proved.

The 13 original PDFs (198,944,687 bytes) retain the prior live release 407039473 /
data/source-pdfs-v1 asset identity checks and manifest linkage; they were not downloaded
again. Archive raw bytes, licenses and source credit remain preserved.
No author program, Sage/native build, new geometry replay or full child gate was
executed in this refresh.
The parent register’s 53 independent tests passed; child semantic comparison covers all
324 rows, and all 170 historical entries match.
This is composed source review, not new full mathematical replay.

Fresh GitHub discovery confirms OWNER jlevy, current child head and parent base, zero
formal reviews, zero inline comments and zero PR comments.
Earlier complete related-issue discovery remains in the audit receipt.
Publisher must repeat that guard.
At this head, Packing required and the exact-side-values job pass; Pages publish and
Pages required fail, while the full run is still active at this review’s observation.
These are distinct outcomes; old failing or passing runs cannot be relabeled.
The foundation must be fixed and requalified before this layer can land.

Review D (correctness review, round 1) · reviewed head
760dd9538a4b2c7c957ce6950d4624d790da3400 · base
6a9b2876d40ebbc4d709ecfecd1bf107e6db95cb.

Reviewer: strong tier, requested gpt-6-astra at xhigh. Trusted channel: independent
review prepared for formal GitHub COMMENT; author jlevy is OWNER in the same
repository. Publication remains coordinator-owned.

Current immediate-layer inventory: 144 changed paths.
Coverage is composed from the explicit prior full audit and the reviewed successor delta
below; unchanged source, byte-only checks and skipped execution are distinguished.

Verdict: no new own-layer mathematical defect found, with foundational landing still
required. Current publication C1 must also be fixed before landing.
The earlier diagnostic child D1 source-proof omission is fixed:
audit_t007_consumers.hosted_paper_sources/paper_source preserves ElMoumni n7/n8/n15
annotations and defects using the manifest, with local/hosted/unrelated controls.
The child’s existing linear-root and full-centre fixes address parent D1/D2 in the
composed tree; they must move into #403 so the foundational PR stands on its own.
The 33 projected finite sides keep unknown optimum/ideal contact questions separate; all
170 historical entries remain identical, including invalid-source flags and
outside-horizon cases.
No root-ordinal, geometry or optimality credit is inferred from source facts.

This is the first formal four-role review prepared for this PR. The earlier 34f7
diagnostic drafts were not published.
It bridges the complete prior 142-path audit (133 exact path bridges plus nine direct
changed-path reads) and the independently retained full source audit.
The preceding 34f7 to 1c102f4f 54-path successor delta has 18 exact main91ca byte
bridges and 36 own changed paths, including parent changes; complete own changes,
callers and records were assessed.
The complete inventory distinguishes direct reads, prior semantic coverage and byte
bridges rather than claiming all unchanged bytes were reread.

The final three-path delta from 1c102f4f to this head was read in full: Pages now
retains the exact archived KKT-results leaf needed by the renderer, the real Git
sparse-checkout contract checks that leaf and neighboring exclusions, and the work plan
distinguishes two reported input downloads from unexecuted solver/search work.
All source data and generated register bytes match 1c102f4f. No new tests or external
input retrieval were executed for this final bridge.

All 170 historical register entries and 291 current entries are identical to the
reviewed earlier register.
The remaining 33 current rows have the same custody-checked finite-rational projection
reviewed and tested in the parent: status, side, lower bound and KKT fields remain
unchanged.
Current totals are 320 exact identities, four numeric rows (29,55,71,105), 170
historical entries and 77 proved cases.
The degree 672 source polynomial, its 673 coefficient strings, all 38 SVG joins, six
article identities and 201 printed occurrences forming 182 coefficient/side pairs retain
their original independent source-audit scope.
The source-stated root ordinal is not independently proved.

The 13 original PDFs (198,944,687 bytes) retain the prior live release 407039473 /
data/source-pdfs-v1 asset identity checks and manifest linkage; they were not downloaded
again. Archive raw bytes, licenses and source credit remain preserved.
No author program, Sage/native build, new geometry replay or full child gate was
executed in this refresh.
The parent register’s 53 independent tests passed; child semantic comparison covers all
324 rows, and all 170 historical entries match.
This is composed source review, not new full mathematical replay.

Fresh GitHub discovery confirms OWNER jlevy, current child head and parent base, zero
formal reviews, zero inline comments and zero PR comments.
Earlier complete related-issue discovery remains in the audit receipt.
Publisher must repeat that guard.
At this head, Packing required and the exact-side-values job pass; Pages publish and
Pages required fail, while the full run is still active at this review’s observation.
These are distinct outcomes; old failing or passing runs cannot be relabeled.
The foundation must be fixed and requalified before this layer can land.

Review F (bounded senior follow-up, round 3) · head
273abfa593d64a01256f8da0059aeae2e4d64186 · immediate base #403 at
3ee60c3e5157a82c0e9b5a48e54fc6c1f57539f5.

Reviewer: strong tier, requested gpt-6-astra at xhigh. Channel: formal GitHub
COMMENT; same-repository OWNER source.

Coverage: inventoried all 177 immediate-layer changed paths, then traced the catalogue
exporter, browser identity resolver, large-polynomial case link, publication/script
limits, selected source/adoption contracts and earlier dispositions.
Read the current committed stack handoff and mathematical review, formal A–E reviews, PR
comments, zero inline threads and all-state linked-issue results.
Scoped range-diff against Review E separates original rebased work from the later
publication/adoption changes.
This is a bounded follow-up, not a fresh complete review of all 177 files, every source
packet or generated artifact, hosted PDF assets, or the full mathematical/native proof
corpus.

Tests run: exact pinned source exports with project Python 3.14.7 and external scratch.
All 17 existing test_exact_catalogue.py controls pass in 29.99s; eight existing
prime-hint/int64 overflow/refusal controls pass in 0.94s. Source exports initially
lacked Git metadata and the KPress import path; supplying pinned Git-object
introspection and the existing exact KPress submodule version resolved those collection
errors. No test assertion or production source was changed.
A maintained exporter run reproduces F1 below.
No complete web build, browser suite, native/Lean or high-degree replay ran here.

Verdict: changes required. F1 is a new user-visible source defect.
F2 and parent integration also prevent final qualification.

F1 — Medium — The n83 complete-coefficient link opens n87.
render_case_pages.py:551
forms #current-polynomial-for--83 from the count when the polynomial exceeds 4,000
characters. The current exact_catalogue.output_files assigns legacy anchors by
polynomial order: current-n83 (degree 672) receives current-polynomial-for--79,
while current-polynomial-for--83 belongs to current-n87 (degree 41). The browser’s
selectHash resolves that latter anchor directly to n87. Thus the n83 case record
promises all 673 coefficients and sends the reader to a different polynomial.
test_case_pages.py:505 currently locks the incorrect literal; the passing catalogue
tests only validate their own mapping.

Fix: link to stable #current-n83 (generally #current-n{n}), or use the
authoritative exported anchor.
Add a producer-to-browser resolution regression proving the case83 link selects n83 and
exposes its 673 coefficient strings.
Preserve the working legacy anchor mapping for old URLs.
Tracked as think-5hv7, a P2 bug under existing stack owner think-6vrv, delegated
to the existing [email protected] owner.

F2 — Medium — The latest full checkpoint has completed with failure.
Run 38387914891 is terminal
at this exact head, replacing the PR body’s pending status.
Eleven separate prerequisite jobs pass, but validate and post-merge-required fail.
The two failed core steps are behavioral shards B/C:
test_without_scripts_the_refresh_still_arrives takes 16.03s and
test_check_requires_a_forwarder_at_every_address_a_page_used_to_have takes 22.42s,
both above the enforced 12s call ceiling.
Core wall is 1,897.21s against 3,600s. The terminal classification reports those wall
refusals, not a functional assertion failure.
Fix: update the PR validation record, diagnose the failing scope under the existing
controlled-timing work, and complete final qualification after F1 and parent
integration. Do not infer a source regression or waive the full gate from green ordinary
PR checks. Timing/qualification remains with think-ht59/think-w0if and think-6vrv.

Prior review assessment: C1’s former 6 MB archive and oversized executable-inline
exceptions are absent.
site_urls.HARD_HTML_LIMIT is 2,000,000; check_site_scripts rejects executable inline
content above 4,096 bytes before classification.
Linked assets are prepared before pages.
The C1 repair has marked dispositions at 7fd2074; old exception-policy criticism
should not be repeated against this source.
E1–E6 have retained marked dispositions and ordinary hosted qualification; this review
does not claim to rerun every original control or newly measure the complete physical
publication.
The missing retained_data.py workflow trigger and consumer declaration are
present. The old modular-hint overflow question is addressed by an operand/degree int64
bound before arithmetic, and the eight selected controls pass.
Parent B1/D1/D2 were independently rechecked in Review J. The documentation relocation
requested after #576 is present in
docs/project/handoffs/handoff-2026-10-10-polynomial-pr-stack.md.

Source/adoption checks: the canonical child retains 324 current plus 297 historical
records; publication projects 337 records.
The 49 current V3/C3 adopted identities and 89 historical native occurrences remain
separate from source-only n106/152/177 roots.
All 324 current side/status/lower projections equal the parent, including 77 proved and
247 open cases. Child n83 supplies the full 673 coefficient strings; its source ordinal
27 remains independently uncounted.
Historical source invalidity and outside-frontier status survive the exporter.
Passing projection tests establish representation fidelity, not geometry, Lean replay,
rigidity or a new global optimum.

Stack boundary: the remote formal stack API confirms stack 447, ID 1893864, with
main → #403 → #435. The child is clean against the exact current parent and contains the
collection/web layer above the register/adoption foundation; that split remains
coherent. The parent now conflicts with fetched main
804eff369193b698d497a9c06847059eb9de0470. Propagate its eventual integration through
the formal stack, then recheck both boundaries; a child-to-parent green run does not
qualify parent-to-main integration.

Current ordinary checks have 32 successes and 26 expected skips, including
Packing and
Pages. Remaining work is F1
repair, F2/current-parent qualification, and final reconciliation of
source-custody/security, correctness and publication-performance review coverage for the
refreshed source. The retained mathematical review exists and remains scoped evidence;
this follow-up adds no new proof-campaign clearance.
No source edits, commits, pushes, stack synchronization, merge, expensive proof replay
or CI rerun was performed.