Add shadowboxing and editorial scar tissue patterns (v2.10.0)

[WesleyDickens]

Fixes #198.

Adds two patterns for drafting-conversation residue: prose that leans on context only the drafting session had, so it reads defensive or digressive to a cold reader.

Pattern 34: Shadowboxing (Defending Against Unraised Objections)

The pattern from #198: meta-level negations about the piece's own aims or the author's intent ("This isn't mainly about X", "I'm not arguing Y") that answer objections nobody in the published text raised.

Two changes from the issue as filed:

• The "would a cold reader arrive at this objection?" test isn't operational: the editor usually can't see the drafting conversation either. Detection instead uses in-text signals — the negation is meta-level (never object-level claims like "the API is not thread-safe"), unattributed, dropped within a sentence, and about a topic that appears nowhere else in the piece.
• The fix guidance resolves a conflict with the skill's claim-preservation rule. "I'm not arguing documentation doesn't matter" embeds a real position; when the piece uses that position, the editor restates it affirmatively instead of deleting it, and deletes only pure not-X clauses. Attributed or genuinely engaged objections stay, with §9 governing their phrasing.

Pattern 35: Editorial Scar Tissue (Phantom Alternatives)

A sibling with the same root cause and a different surface: mid-argument, the text rebuts a "tempting" alternative no reader would consider, drops it, then does it again later on an unrelated tangent. The alternative is usually the model's own corrected mistake, recycled as a strawman — a scar from the drafting conversation, where the option was live until a human killed it. Tells: attributed to no one, appears nowhere else in the piece, dismissed in a clause or two.

The general test, which covers both patterns: if you can explain which previous edit caused a sentence to exist, rather than what new information it contributes, it's scar tissue.

Together with §30 (diff-anchored writing), these cover the three ways text leans on invisible history: commit history, drafting objections, and corrected mistakes.

False-positive guards

"This isn't about X" in the wild is usually legitimate, so both patterns come with guards in "What NOT to flag": scoping statements ("This guide does not cover Windows"), legal and safety disclaimers, corrections of misconceptions readers actually hold, attributed or engaged objections, replies and FAQs that answer someone by design, voice-bearing asides, and alternatives a reader would actually reach for (design docs weighing real options, tutorials warning against genuinely tempting mistakes, essays that steelman).

Related changes

• §24 (Excessive Hedging) gains watch phrases and the accumulation dynamic: iterative editing stacks fairness clauses ("to be fair", "it's also possible") until every conclusion reads as negotiated. A claim earns one honest qualifier at most; a caveat that exists only because an earlier draft overreached gets cut with the overreach.
• The revision step in Process and Output now says to rewrite from the point rather than patch the flagged phrase. Sentence-by-sentence patching is how scar tissue forms, and the skill's own draft → audit → final loop is not exempt.
• Maintenance contract: README heading and table updated to 35 patterns, AGENTS.md count updated, version bumped to 2.10.0 in all three synced places, and the validator's pattern ranges extended. python3 scripts/validate-package.py passes; SKILL.md is at 438 of the 500-line budget.