Two-channel fidelity: harmonised semantic construct vocabulary plus a quarantined residue channel

[Mearman]

Each codec's README documents what its format carries that the content model drops. This issue replaces those implicit, per-repo caveats with one explicit, two-channel model -- and defines "full fidelity" as three testable tiers.

Channel 1 -- semantic constructs as first-class nodes, in one harmonised vocabulary (no format-specific node kinds ever):

• contentControl -- docx SDTs, PDF form widgets: a typed container of content.
• field -- docx fldChar/instrText ranges, ODF field masters: instruction plus cached result, extent as containment rather than marker pairs.
• anchor -- docx bookmarks, ODF text:bookmark, markdown footnotes: a named extent wrapping its content.
• link -- hyperlinks and PDF link annotations: target plus wrapped runs.
• provenance -- tracked changes (w:ins/w:del, text:tracked-changes): author/date wrappers around content.

A construct with no cross-format analogue (docx SDT with no ODF counterpart, ODF xforms with no OOXML one) does not get a bespoke node: it degrades to the nearest generic node with the specifics in the residue channel, decided once and documented. The tree is organised by what documents mean; the residue channel is how a format serialises.

Channel 2 -- quarantined residue: a per-node and package-level source: { format, xml } field, schema-validated as opaque text, never semantically interpreted, carrying what has no cross-format meaning (mc:AlternateContent internals, custom XML stores, proofing marks, raw HTML blocks in markdown). This generalises existing practice: starMath beside canonical MathML on ContentFormula, styleId's opaque source names, sourcePath traceability.

The three tiers, made precise:

• Semantic fidelity: every translatable construct survives as a first-class node, so cross-format conversion loses no meaning.
• Restorable fidelity: same-format re-emission rebuilds from semantic nodes plus residue, verified by round-trip tests over the corpus.
• Byte fidelity: the edit model's job (live views over raw XML), for any format, forever. No semantic pivot achieves byte identity; the model does not pretend otherwise.

The construct verdicts (semantic / residue / derivable -- derivable meaning recomputable and dropped without loss, e.g. pivot caches, statistics parts) are recorded per format in the four inventory issues: ooxml.js, odf.js, markdown-codec, pdf-codec. Those lists feed this vocabulary; this issue blocks on them.

[Mearman]

Two reconciliation decisions to fold into this issue's design, both surfaced by the implementation review:

• Links: ContentRun.hyperlink (src/content.ts) is today's link vocabulary -- markdown-codec carries all links through it, and documents.js's from-package derives its LayoutLink items from it. The link node kind does not replace it: run.hyperlink stays the run-level encoding, and link nodes are reserved for block-scoped or annotated link extents that a flat run field cannot express. Unstated, this would re-litigate markdown-codec's both-direction z.codec validation and the frames-to-items inverse later.
• Footnotes do not fit anchor as written: a footnote is a marker at the reference site plus a definition at the document tail -- not one contiguous extent. The definition becomes an entry in the package-level definitions-table facility riding Promote DocumentPackage to the single hierarchical tree: structure, layout, and content in one artefact #20's major (same mechanism as the styles table of Package-local referenced styles and group inheritance: compression-only, materialised on flatten #21); the marker is the anchor. Also for the inventory's attention: markdown footnotes have zero support today (nothing is carried), so this is new reading work, not a mapping exercise.