docs(architecture): retarget the engine documentation to 2.x

[DemchaAV]

Why

The architecture pages still describe the Entity/ECS engine removed in 2.0. A reader following them reaches for RenderStream, RenderPassSession, TextComponent, EntityManager, ParentComponent, EntityBounds, ParentContainerUpdater, LayoutTraversalContext, ComputedPosition, Breakable, and the packages engine.core / engine.pagination — none of which exist on any live branch. pagination-ordering.md is the worst case: its entire "Current implementation shape" section documents a priority-based topological walk that no longer runs. The release runbook calls itself the 1.x checklist and points every command at aggregator/pom.xml and -pl ., neither of which is the 2.x layout.

What changed

• overview.md drops the ECS responsibilities section, names both fixed-layout backends in the pipeline and the package list, and gains a PPTX leg in the diagram. That diagram also had the DOCX edge on the wrong node: DocxSemanticBackend.export takes a DocumentGraph, so it branches off the semantic tree, not off the resolved layout — which is exactly why it cannot reproduce fixed-layout geometry.
• package-map.md rebuilds the engine table from the live package list (components / debug / font / measurement / render / text), splits the row that claimed render-docx and render-pptx were both semantic exporters, and adds document.backend.fixed.pptx.
• lifecycle.md loses its v1.2 title, gains the PPTX render path, and states chrome coverage per backend instead of calling metadata and watermarks PDF-only.
• pagination-ordering.md describes the mechanism that actually runs — a single sequential LayoutCompiler pass whose emission order is render order, with descendants-first a property of the recursion rather than of a sort. Its debugging steps now name the live pagination log events and layoutSnapshot().
• release-process.md becomes the 2.x checklist: the whole-reactor gate, the 13-pom lockstep list taken from the script's own array, the real staging allow-list including the per-module READMEs, and a branch-flow section describing the standing roles rather than the completed 2.0 transition. It also records two traps: a gate command ending in a pipe reports the pipe's status and hides BUILD FAILURE, and after a version bump every goal resolving train modules from ~/.m2 needs an install first.

BlockText and ContentSize are still live and were left alone, as was the logger category com.demcha.compose.engine.pagination. Two 1.x references are kept deliberately: the backport callout (the 1.x branch genuinely uses aggregator/pom.xml) and the dated v1.6.5 lessons-learned entry.

Verification

• ./mvnw -B -ntp clean verify — BUILD SUCCESS, 1508 tests, 0 failures.
• A sweep for every dead identifier listed above returns nothing under docs/architecture/.
• Every internal link added resolves on disk.