Plan FDX0 evidence-grade IR for datasheet training

[akiselev]

Summary

Adds Ferrodoc's upstream FDX0 implementation plan for evidence-grade DocumentIR suitable for cheap downstream datasheet extraction/training.

FDX0 keeps Ferrodoc domain-neutral while adding the generic capabilities the downstream Datasheet compiler needs:

• explicit geometry-quality semantics rather than pretending page-sized text is precise;
• positioned native PDF text research/implementation path using the existing Rust stack first;
• evidence-bearing table cells with source-span references;
• a deterministic born-digital table recognizer with explicit abstention;
• table.recognize integration through the existing engine abstraction;
• selective page/region refinement as a new immutable evidence generation;
• optional learned table-engine evaluation only after the deterministic baseline is measured;
• storage/accuracy/resource benchmarks and acceptance gates.

docs/fdx/README.md maps the complete cross-repository FDX0-FDX6 program while making ownership explicit: Ferrodoc owns only FDX0/generic document evidence.

Boundaries

FDX0 does not add electronics predicates, part-number semantics, source APIs, Foundry lake types, unit semantics, or the downstream model.

Companion PRs

• Foundry integration: https://github.com/akiselev/foundry/pull/5
• Datasheet-cli FDX1-FDX6: Plan FDX1-FDX6 IR-native datasheet extraction datasheet-cli#2

These three PRs define one coordinated implementation sequence; Ferrodoc implementation remains independently mergeable as the upstream FDX0 dependency.

Validation

Planning/docs-only PR. Branch comparison against master is clean and adds:

• docs/fdx/README.md
• docs/fdx/FDX0_EVIDENCE_GRADE_IR.md

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 33eabd0fb5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Qualify region targets with their page

When a valid imported DocumentIR reuses a RegionId on different pages, this target cannot identify which region to refine: Document::validate checks region-ID uniqueness only inside each validate_page call (crates/ferrodoc-ir/src/lib.rs), not across the document. This makes the promised region-only execution ambiguous and can refine multiple regions or the wrong page; include the containing PageId/page index in a region target, or make region IDs document-wide unique and validate that invariant.

Useful? React with 👍 / 👎.