Expand and harden the correspondence engine

[jcushman]

This PR advances binoc's correspondence-first engine with a coordinated set of dataset configuration, tabular analysis, rendering, performance, plugin catalog, and test-infrastructure changes.

• Makes dataset.paths the sole per-path and per-table configuration surface, with descriptor-derived dispatch, content_type and forced-rule overrides, CSV dialect and shape facets, nested JSON records_path, row identity, structured-document node identity, and explicit validation for invalid facet/rule combinations.
• Keeps dataset.defaults.row_identity as the global default and uses presence-aware paths[].row_identity patches so named and positional selectors and null/duplicate policies override correctly. Removed the obsolete dataset.tables model, flat row-identity aliases, selector merge/fallback code, and coexistence tests.
• Expands tabular edit-list rewriting with column rename and reorder composition, type-only column changes, sorted row alignment, reduced-precision and suppression claims, row-addition consolidation, and a documented single-pass rule order.
• Adds inference disclosure and conservative guardrails for auto-detected row keys, dialects, high-churn positional comparisons, and uncertain row correspondence.
• Improves CSV parsing for declared and inferred delimiters, quoted fields, decimal-comma data, banner/header lines, extensionless inputs, and false-positive resistance.
• Improves large-table performance with shared artifact decoding, a faster keyed writer, configurable in-memory thresholds, bounded streaming keyed comparisons, recorded performance baselines, and user-facing tuning documentation.
• Adds bounded binary CDC range localization, deterministic source-materialized binary fixtures, UTF-8 boundary-safe content sniffing, and extensionless text/binary classification.
• Streams archive and gzip expansion with per-entry and total decompression caps, incremental hashing, cleanup of partial outputs, and clear fallback diagnostics.
• Tightens semantic writer ownership so parsed tabular or structured content is not also reported as raw text, while parser metadata and structural container facts still compose independently.
• Makes projection and Markdown rendering more rule-driven through generic structured summaries, producer-backed content_summary facts for modified moves, bounded and round-robin examples, configurable verbosity/grouping, deduplicated details, humanized counts, and renderer-ignorance invariants. Removed the dead tabular_summary renderer channel.
• Adds offline feedback-report bundles, reproducible snapshot capture/reference modes, and extract commands in configuration diagnostics.
• Aligns the plugin catalog and Python packaging with the fat-binoc distribution. Removed the contradictory binoc[all] extra and the ignored registry= argument from binoc.diff.
• Extends SDK and engine support for projection annotations, package identifiers, per-item dispatch hints, artifact decode caching, fallback writer composition, and source-aware extract hints without adding format knowledge to core.
• Adds and strengthens behavioral test vectors, structural invariants, description-cost baselines, stdlib/plugin lints, generated gallery/replay documentation, and the native renderer ABI canary.
• Updates active docs and ADR statuses to make the per-path and fat-wheel decisions authoritative while preserving historical ADR text.

Validation:

• just fmt
• just check
• just test
• just lint
• UV_NO_CACHE=1 just docs
• git diff --check main...correspondence-hardening

(Auto generated pull request)