docs: freeze Infrastructure Atlas architecture gate

[Joncallim]

Documentation-only architecture proposal for #250/#251 after repeated repo-grounded orthogonal review. No production behavior changes.

Files:

• docs/architecture/INFRASTRUCTURE_ATLAS.md — implementation contract and issue spine.
• docs/architecture/INFRASTRUCTURE_ATLAS_REVIEW.md — hostile review ledger, now extended through a second convergence sweep.

Second-sweep hardening added:

• port/exposure semantics may not be inferred from opaque ContainerRecord.ports: string[];
• semantic containment and presentation-only lanes are separate model/visual concepts;
• inspection scope is not automatically host containment for external/runtime context nodes;
• Docker/service/runtime representations may only be correlated/merged with proven canonical identity, never label similarity;
• conflicting correlated facts need explicit source ownership/diagnostics, never last-write-wins;
• Atlas must consume the existing generation/provenance/modelRevision-coherent model path rather than screen-level piecemeal topology fetches;
• camera stability is now an independent contract so auto-fit cannot defeat logical layout stability;
• global density/theme settings cannot re-anchor/reorder the Atlas;
• canvas, directory/text alternative and inspector share one semantic selection/order authority;
• aggregate/count language must preserve resolved/unresolved/ambiguous populations;
• accessibility tests require deterministic keyboard order independent of SVG paint/insertion order.

Earlier key corrections remain:

• current ContainerRecord has no Compose project identity: V1 cannot assume/infer project grouping;
• semantic Atlas output excludes ephemeral publication revision/time from exact deterministic goldens;
• collided visible occurrences need canonical presentation occurrence handling, never array-index routing identity;
• structural stability is defined by mutation class;
• Overview is orientation-first, not all-edges-first;
• health, freshness, Findings attention and ambiguity remain orthogonal;
• zero graphics/layout dependency is the baseline to beat;
• global O(n²) layout work is unacceptable at the 250-object stress gate;
• semantic/layout goldens, geometry tests and screenshots have separate authorities.

The second hostile convergence sweep produced no additional independent P1/P2 architecture class after these refinements. Remaining choices are deliberately authority/fixture driven (#259/#262/#252/#253/#257/#268/#261/#269), not implementation taste.

Keep draft until the maintainer accepts the architecture gate; no production renderer should start before #251 closes.