Teach the design-intent-verb rule, then warm Part III under it

[dcellison]

Two commits, in the order the work was ruled to happen: the voice guide learns the design-intent-verb rule first, then Part III's warmth revision follows the updated guide.

Voice guide update

documents/reference/voice_for_models.md gains the rule that governed the Part I and Part II restorations but lived only in PR bodies until now:

• A new anti-pattern row, "Effect verb under an image": imagery credited with changing the speaker is zero-tolerance; the image hangs on a design-intent verb (built to, asks, offers, invites) and the changing is left to the speaker.
• A new gallery pair from the real Part I history: the original manuscript's "beginner's mind" effect claim beside the merged word-order passage, with the lesson stated: the cure for a false effect claim is rarely deleting the image; it is changing the verb.

Part III warmth revision

Four of the eleven files change; eight lines. The register work, under the hedge budget and the new rule:

• The chapter 7 introduction loses its waiver-shaped ending ("what happens there is still the speaker's work," the same possessive tail the Part II counting pass flagged as a stamp) and gains two things the original manuscript had on false verbs: the stream over smooth stones as the texture the three rules give, and the vista, stated honestly, that nineteen sounds and three rules put every Phi word within the mouth's reach, including words nobody has coined yet. Chapter 7's one full-strength waiver is now the digraph file's existing close ("producing /ɸ/ does not certify anyone's state of mind"), which was already the best sentence in the family. A stray self-reference ("Chapter 7 gives...") inside chapter 7 becomes "This chapter."
• The golden rule file gains the manuscript's pearl image, recast physically: each syllable a pearl on the word's thread, the boundary the knot that keeps two pearls from wearing into one bead.
• The syllable structure file's deflating last line (ending the file on "remains a question for listening evidence") keeps its honesty but ends on the offer: a Phi word's joints show through its spelling, which is more than most languages let a beginner say.
• The writing file closes Part III with a threshold, matching the thresholds Parts I and II now have: the opening promise recalled and fulfilled, and the turn to Part IV arriving in words the reader can already say.

Seven files are untouched: the vowel, consonant, and IPA reference pages are clean reference register, and the stress, transfers, and exercises files already end well ("one stray glide is an event, not a diagnosis").

Judgment calls to review

• "Nineteen sounds, three rules" appears in both the chapter 7 introduction and the Part III threshold. This is a deliberate opening-promise, closing-fulfillment bookend on the Part I pattern, not an accidental migration; if it reads as a tic, cutting the threshold's restatement is the right fix.
• The original manuscript's claim that Phi's sounds are accessible because "most humans already produce" them was not restored; the current files' honesty about /ɸ/, /θ/, and /ʍ/ being genuinely hard is better and stays.

Audit

The full voice_for_models.md protocol ran in order, including the row-by-row counting pass on the diff. Named findings, corrected before commit: the chapter 7 introduction's old ending carried the exact "speaker's work" possessive tail the Part II pass had just flagged, and converting it removed a corpus-level stamp; the vista draft initially reused "instrument" from Part I's laboratory close and was reworded; the pearl sentence was recast to avoid placing an American spelling beside this file's existing British "neighbours." Counting results: all hard-zero rows at zero, one earned triple (the promise's literal content), one aphorism-class image per document, waiver count now one full-strength in chapter 7 and zero in chapter 8, whose remaining "cannot" sentences are grammar facts. Validator: 1275 entries, 0 errors, 0 warnings.