Prose-gate unpaired additions and deletions

[mmcky]

Closes #14 — the dominant pair-precision cause measured by the three-edition adjudication (QuantEcon/project-translation#53).

An unpaired line with no target-script content — code, maths, markup, or .translate frontmatter arriving or leaving — now routes to code-or-markup instead of addition/deletion, mirroring what categorise() already does for non-prose paired lines. The omission taxonomy therefore counts prose only, which is what the adjudication showed it was failing to do (addition precision 4/13 / 8/18 / 1/20 across intro / python / programming, the misses almost entirely code and metadata rows).

As #14 recommended, non-prose lines are re-routed rather than dropped, so the total pair count is unchanged — the regression test pins the corpus pair count against the 2026-08-04 study, and a category-only shift keeps it green by construction (verified: full suite passes including the pinned regression).

Real-corpus effect at programming.zh-cn @ b727528, the pathological case: addition 103 → 30, deletion 80 → 38, code-or-markup 25 → 140, total 255 unchanged; composition and engine_strata byte-identical. The direction and scale match the adjudication sample, where 19 of 20 sampled additions were non-prose.

The synthetic fixture gains one purely additive and one purely deletive editor-shaped commit, so the routing is asserted without depending on git's hunk alignment or the similarity pairing (which cross-pairs freely inside mixed hunks — the separate, smaller artefact #14 leaves out of scope). docs/method.md's line-pairing known-limit now states the rule.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR fixes a precision defect in the omission taxonomy by applying the same “prose gate” used for paired lines to unpaired additions/deletions during scanning, so that non-prose (code/maths/markup/metadata) lines no longer inflate addition/deletion.

Changes:

• Route unpaired added/deleted lines to code-or-markup when they contain no target-script characters, otherwise keep addition/deletion.
• Add a synthetic baseline test that asserts the new routing for a purely-additive and purely-deletive edit shape.
• Update method documentation and the changelog to describe the rule and its intended effect (category-only shift; total pair count unchanged).

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
src/textstrata/scan.py	Prose-gates unpaired additions/deletions and assigns taxonomy via CATEGORY_MAP to keep non-prose edits out of omission buckets.
tests/test_baseline.py	Extends the synthetic repo fixture and adds a focused test asserting unpaired-line routing and taxonomy for code lines vs prose.
docs/method.md	Documents the known-limit rule that unpaired non-prose lines are categorised as code-or-markup.
CHANGELOG.md	Records the behavioural change and clarifies that totals are preserved (only categories shift).

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.