Review of the second Knowledge Lineage overview

In short. The inward overview, thread and evidence levels preserve the distinction between references, documented inquiry choices and empirical findings. The frozen population and all eight selected relationships check out. Fix the node evidence pinpoints, make the download result's population boundary visible before disclosure, and retain legible drawing text between phone and desktop widths. The desktop return arrow was wrong in the inspected proofs; its source was corrected during this review and needs a regenerated proof.

10 November 2026. Fresh OpenAI Codex agent context; exact host model identifier and attributable cost unavailable. This is an agent source/interface review contribution, not independent confirmation, a human acceptance test, or a release review of the original research. No external pages, credentials or network workarounds were used. Only this review note was written; no implementation files or commits were changed by the reviewer.

Actionable findings

P2 — Some “Retained basis” links do not contain the observation they purport to support. overview-history.json's advisory passage is what-moves-a-pin.md L28, about open-range images, whereas its summary concerns pin moves after advisories and the failure of the burst to recur. The relevant frozen passages are L38 (six of thirteen observed moves within 28 hours), L14 (the later qualification) and L46 (timing does not establish the reason). The images passage is only L28's recipe mechanism; the observed version preceding the lockfile is L30 and the no-run boundary is L54. Similarly, four omits its no-burst and request-only passages (L34/L56 and L62), and notes omits the two-library concentration and later uncertainty (L58 and L14). Use multiple bounded passages or an explicit span per node, with links for the consequential qualification as well as the original result. Preserve the full basis in the export. The summaries themselves are broadly supported by the articles; this is a pinpoint defect, not a finding that the observations were invented.
P2 — The fast-download branch hides the boundary that keeps its visible story accurate. The drawing's “A tested result, with a boundary” does not tell a closed-disclosure reader what that boundary is. The introduction still calls downloads fast-moving generally. The actual result is that most releases reached half of downloads within two days in 26 of 37 heavily downloaded libraries; the same result was not the rule in the lower-rank samples. Keep that population and the lower-rank qualification beside the drawing or in the visible caption. In the expanded summary, replace the undefined “two-day pattern held” with the actual majority-of-releases measure. Only 127 of 424 releases met the threshold when counted by release, so the per-library wording matters. Do not imply that low download share means slow uptake. The current summary's fetch/install boundary should remain.
P2 — The drawing becomes tiny just above the phone breakpoint. lineage.css selects the 1040-unit-wide drawing above 680px, while the drawing's titles are 15px and its status text is 12px at natural size. A 681–768px viewport leaves roughly 649–736px after the normal gutters, reducing these to about 9–11px and 7–8.5px. This is a source-level sizing consequence, not a browser measurement. Switch to the narrow layout until the wide drawing can retain readable type, or provide an intermediate layout. Check an intermediate-width static proof after changing it. The supplied 390px and 1100px proofs alone do not exercise this interval.
P2, source repair observed — The desktop return-to-downloads arrow pointed out of its target. The original draw_overview.py path M245 554 H1025 V124 H1028 terminated to the right of the pace box, pointing away from it. The defect is visible in the inspected 1100px light and dark proofs; the phone arrow enters the box correctly. During this review the source became M245 554 H1032 V210 H910 V171, which approaches the target from below. Regenerate the SVG and proofs and verify that the corrected arrow visibly enters pace. No new inquiry relation is needed.

What checked out

Independently counted the frozen tree at 8af0cb026037a8ee63456900978b3c44473c49cd: 3 investigation declarations, 49 study directories with top-level Markdown leads, 18 released top-level articles, and 3 released top-level essays. All 21 publication headers say released. The extraction excludes the 17 nested article short forms and 79 study checkpoint documents from node membership. There are no draft top-level publications at this frozen boundary, so draft exclusion is established by the filter rather than an observed excluded draft.
All 73 catalogue nodes are reachable through the three investigation groups plus “Other work”. The ten unassigned records are LH000, LH003, LH004, LH016, LH030, LH032, LH035, LH037, LH039 and LH040. Group overlap does not inflate the headline population. No unresolved internal Markdown target or reference-style link syntax was found in the retained lead/brief inputs.
The 228 generated references have a separate explicit reference status and a scope that rejects support, influence and causation. Declared membership is distinguishable from an editorial history relation. The coverage text correctly says that study-record existence is not reviewed knowledge, and that this is an inward catalogue rather than a collected external research landscape.
Read the frozen released passages for the selected eight pieces, including the important corrections and later-evidence notices. Each history edge's exact retained passage and SHA-256 matched its frozen file. Direction is earlier reading to later reading, with the predicate describing the later reading's treatment of the earlier one.

Relation	Frozen primary support	Assessment
Projects → downloads, extends	`what-a-download-shows.md` L10	Explicitly extends the named nine-project reading.
Downloads → images, extends	`what-an-image-holds.md` L10, L48	Moves from fetched files to packaged installations without claiming a run.
Images → advisory reading, tests and branches	`what-moves-a-pin.md` L10, L56–L58	Tests the two-pace reading on a later-release sample and asks a conditional advisory question.
Advisory reading → four libraries, revises	`pins-move-over-weeks.md` L20, L56	Revises the suggested general rule; it does not retract the one-library counts.
Four libraries → release notes, narrows	`pins-move-over-weeks.md` L66; `fixes-that-said-so.md` L10, L66	Narrows the release/advisory question to what a fix's own notes disclose.
Release notes → thirty libraries, widens and revises	`the-lead-was-a-few-libraries.md` L12–L14, L26, L74	Unsettled effect, not a null effect; twelve fixes overlap earlier readings, so no independent replication is established.
Downloads → wider pace test	`two-days-for-most.md` L13, L71	Explicitly tests the one-library case more widely.
Thirty libraries → wider pace test, returns to a question	`investigations/software-updates.md` L63	The investigation records this change of research direction. It is not a causal result about software.

The new summaries do not resurrect “notes made no difference”, shares standing still after their first days, a general advisory burst, installation as execution, or download share as deployment uptake. The current-account link explicitly includes later notices. The later evidence behind the lower-rank boundary remains one disclosure too deep, as finding 2 explains.

Interface and verification limits

Inspected all four supplied overview proofs: 390px and 1100px, light and dark. They are readable static drawings at those sizes; the 390px layout keeps all eight nodes and the side branch. The desktop proof has an empty “Take a closer look” area even though both links are present in generated HTML. That is a proof-rendering limitation to resolve before treating the image as a faithful page proof, not evidence that the browser hides those links.

Read the three Nunjucks views and CSS, and checked the existing generated HTML's local anchor destinations: 140 overview, 55 thread and 26 evidence links; no missing local file or fragment. Ordinary links, native details/summary, all catalogue members and the text history exist in HTML without client rendering. The embedded SVG is an image: clicking it first opens the drawing, whose boxes then link to articles. Say this directly in the caption/alt text; “each box opens an article” currently describes the opened SVG rather than the embedded image.

No browser launch, keyboard navigation, assistive-technology behavior, unfamiliar-human comprehension test, or subdirectory build was performed in this review. The static proof process changes CSS and disclosure visibility for a print renderer, so it does not establish browser behavior. No old 8/8 score is evidence for this iteration. A single bounded reread after implementation fixes is appropriate; remaining actual-browser and human acceptance limits should stay explicit.

One bounded final reread

10 November 2026, after the driving session's fixes. The four findings above are resolved in the inspected source and generated data. No further blocking source or claim issue was found. The original findings are retained as the review history; they do not describe the final implementation's current state.

Finding	Resolution checked
Evidence pinpoints	All eight nodes now carry labelled `evidence_spans`, 18 in total. Independently matched every passage, recorded line range and SHA-256 to the frozen primary article. The images result is now L30 with the no-run limit at L54; the advisory result is L38 with its later qualification at L14. The four-library and fourteen-library nodes now retain the missing request-only, concentration and later-uncertainty passages. The projects limit is the exact bounded-scope conclusion at L50. The template exposes these as separately labelled evidence links.
Download population boundary	The summary now states that most releases reached half the downloads within two days in 26 of 37 heavily downloaded libraries. The visible figure caption says that the later lower-rank samples did not show the same pattern, and the drawing's status names the top-library population. The fetch/install distinction and unsettled notes result remain.
Intermediate-width text	The narrow drawing now remains selected through 1071px, allowing approximately the wide drawing's natural 1040px plus the normal gutters before switching. The narrow drawing is capped at 360px rather than expanded across the tablet. Its rotated side annotation was replaced with a horizontal footer label. This resolves the identified source-level shrinking interval; no browser sizing measurement is claimed.
Desktop return arrow	The generated wide SVG contains `M245 554 H1032 V210 H910 V171`, approaching the pace box from below. Both SVGs retain eight relationship arrows and eight document links. This resolves the wrong endpoint in the SVG data; the final visual regeneration remains the driving session's check.

The overview now embeds the SVGs inline, so the eight boxes in each layout link directly to their articles. Both drawings have distinct title, description and marker IDs (history-wide-* and history-phone-*), their label/marker references resolve within their respective SVGs, and the inspected generated overview HTML has no duplicate IDs. The HTML text equivalent remains available through native disclosure. The generated HTML encountered during this reread still carried the previous projects-limit link at L40 while the current exported evidence correctly records L50; the announced final rebuild must bring that output in line with the source. This is a pending regeneration step, not an unresolved evidence annotation.

The static-proof source now explicitly restores the trail-choice grid and substitutes inline SVG palette variables. The next proof regeneration can therefore check the desktop trail doors and the corrected arrow. I did not rerun that renderer or inspect a newly generated proof in this final reread. The population remains 3/49/18/3, with 228 explicit references and ten unassigned records. No new research-benefit claim or superseded result was introduced by the fixes.

This completes the single bounded reread. Actual browser, keyboard, assistive-technology and unfamiliar-human acceptance remain open exactly as before; the inline SVG's accessibility behavior has not been established by the source/ID check.