LAB-005: Probe generator and the first Word measurement study

Context

The measurement model (how Word turns font units into advances, where it rounds, how it computes line heights, when it kerns) is the foundation of line-break agreement. This first study establishes the reverse-engineering method and answers the three most basic questions, WBN-0001 to WBN-0003.

Objective

A reusable probe generator, three confirmed Word Behavior Notes, and a prototype implementation that reproduces Word’s line breaks on the probe set with Word’s own fonts.

Scope

In scope

• Probe generator (Rust): writes minimal valid .docx files from parameters (font, size, text pattern, paragraph width via margins or indents, spacing, kerning flag, ligature settings, compatibility mode), recording parameters in the corpus manifest as T0 documents.
• WBN-0001 (advances and rounding): probes varying string content and length, font size (for example 8–72 pt including fractional half-points), and width, for the most common fonts; observe where Word breaks lines; infer how advances are scaled, rounded and accumulated.
• WBN-0002 (line height): probes for single, multiple (“auto”), exact and at-least spacing across fonts and sizes; infer which font metrics Word uses and how it rounds.
• WBN-0003 (kerning and ligatures defaults): when Word applies kerning and ligatures by default and how w:kern and w14:ligatures change it.
• Notes written with the template, including probes, observations, rules and confidence.
• Prototype: implement the rules in the CORE-003 pipeline (or an early bayan-text), run it on the reference machine with Word’s own fonts (algorithm fidelity), and measure agreement on the probes.
• Add the probes to the T0 corpus with their ground truth.

Out of scope

Paragraph features beyond spacing (tabs, justification: Phase 1 LAB-102).

Deliverables

Probe generator, three notes in docs/specs/word-behavior/, prototype rules, study report.

Acceptance criteria

• AC-1 WBN-0001, WBN-0002 and WBN-0003 are written with High confidence, or with an explicit account of what remains unexplained.
• AC-2 With Word’s own fonts on the reference machine, the prototype reproduces Word’s line breaks on at least 99% of the probe set and line heights exactly.
• AC-3 Probes and ground truth are stored in T0 and runnable as regression tests.
• AC-4 No Microsoft font data leaves the reference machine.

Verification

Study report with metric tables; tests runnable in CI with the open-font variants of the probes.

Escalate if

Observations cannot be explained by any consistent rule after systematic bisection; describe the anomalies so the planner can decide how to proceed.