fix: restore the complete engineering story and sketchbook README rendering

README.org

@@ -84,14 +84,16 @@ POO/Core/Flow/module and contract mechanisms; it does not define another package
 loader or model runtime.
 
 * Engineering workflow
-This diagram expresses the intended integration and its current stop. The actual
+The LLM agent chooses investigations and proposes Library extensions; ASCENT
+exposes evidence and changed premises; Aitia retains admission and authorization.
+These relationships form a feedback loop, not a fixed sequence of agent actions. The actual
 [[file:scenarios/project-formation/func.ss][capture preparation source]] still returns
 =capture-blocked= and =capture = #f= for DE-CASE-01. A diagram or successful render
 does not issue a capture, accepted query view, review result or authorization.
 
-[[file:docs/90-99-views/90-architecture-views/90.02-agent-evidence-loop.svg]]
+[[file:docs/90-99-views/90-architecture-views/90.02-agent-evidence-loop.png]]
 
-[[file:docs/90-99-views/90-architecture-views/90.02-agent-evidence-loop.d2][Editable D2]] · [[file:docs/10-19-design/10-architecture/10.31-architecture-document-and-decision-contracts.org][Architecture/document/decision contracts]]
+[[file:docs/90-99-views/90-architecture-views/90.02-agent-evidence-loop.d2][Editable D2]] · [[file:docs/90-99-views/90-architecture-views/90.02-agent-evidence-loop.svg][SVG export]] · [[file:docs/10-19-design/10-architecture/10.31-architecture-document-and-decision-contracts.org][Architecture/document/decision contracts]]
 
 Methods and content checks must keep their actual inputs, result semantics,
 source/revision associations and open findings. ASCENT reasoning consumes explicit

docs/10-19-design/10-architecture/10.146-taste-selection-and-reuse.org

@@ -109,3 +109,32 @@ HTML/SVG third-party font notice sets are preserved, while the old project owner
 notice is absent from current tracked files. The concurrency test body is byte-for-
 byte unchanged after its two header lines. REUSE now reports only
 LGPL-2.1-or-later and OFL-1.1 as used licenses, with no Apache requirement.
+
+* Engineering overview hotfix after PR #2 squash
+The main baseline is 2ec95bf. The original 90.02 model in 253a926 has seven
+subjects and ten relations, including the LLM agent, Library, ASCENT, progressive
+evidence, investigation, observations and Aitia. The migration had replaced that
+model with a different capture-preparation diagram. Restore the original semantic
+closure; keep capture-blocked as a current-source footnote rather than the story.
+
+A flat ELK drawing preserves edges but provides poor reading hierarchy. This
+project composition now lets source-declared Questions group a purpose summary,
+a compact responsibility diagram, investigation/evidence explanations, knowledge
+and operational feedback, and a takeaway. Subjects and associations bind to the
+same SDK model. The project compositor reuses composeHeading, composeDiagram,
+modelDiagram and selected theme resources; it adds no model or layout solver.
+
+All visible text remains SVG annotations or scene text, so the SVG and PNG retain
+the same story content. Unknown Question, subject, caption, relation and color
+references refuse. Required Design associations cover the complete original loop;
+hiding a narrative association must fail the actual export. The standard generation entry enforces
+12 CSS-pixel minimum text and a bounded portrait aspect ratio for this Case; these
+are Case acceptance constraints, not a universal aesthetic score.
+
+The GitHub raw SVG endpoint sends default-src 'none'; style-src 'unsafe-inline';
+sandbox. A real browser inspection reproduced fallback type there despite embedded
+Kalam fonts; a local data-image probe had misleadingly loaded them successfully.
+README therefore displays the same renderer's PNG and retains an SVG export link.
+The display test binds the actual PNG receipt, observed subjects/relations, measured
+readability and captured source hashes; checking
+only the D2 theme selection did not prevent the original display/content regression.