fix(companion): adopt flatter full speech mouth from Avatar 0.10.1

[DoodleBears]

Closes #1241 · Spec #1135 · Owner PR #14 · Agent claim · Task codex/local/01a11734-bb09-7f01-9d45-adc0f5df6269

Why the change

Window Companions use a flatter full speech opening in place of the rejected circular O mouth while keeping the accepted half-open mouth and speaking rhythm.

Special things to note

• Merge risk: two-way door, small blast radius — revert the rendering option and package pin; only Companion speech opts into version2, with no stored recipe migration. Review default/portrait compatibility and all pre-rendered head turns.
• Prerequisite feat(client): polish companion sound, motion and message reading #1180 is merged; this PR now targets main. Human explicitly authorized publishing/adopting formal pixel-avatar0.10.1; release workflow passes and npm confirms the version. This uses no local tarball or patched distribution.
• Human accepted the formal mouth recording on2026-10-09. Historical captures keep their original revision attribution; broader Window Companion: opt-in text babble and spring-based drag tilt #1173 lifecycle and Window Companion: measure frame and heap costs and agree performance budgets #1167 performance qualification remain separate.

Change outline

Current merge checkpoint

Head 1418c0ca incorporates prerequisite 43dae153, including current main8c2103de, without changing the accepted mouth consumption or speech rhythm. Both bilingual additive conflicts preserve the mouth and floor corrections; released history remains immutable. Local typecheck,13focused renderer/Companion regression tests and bilingual ledgers pass. Standards and Spec final integration reviews each find0defects. Complete CI38071342721 passes at1418c0ca. Final squash-history synchronization38571a9c restores the sole additive mouth entry without changing that tested tree: both have Git tree618217329cc4a7e997e6322d0fe5d4ef559e5354. Squashed main88ced4d1 is tree-identical to the tested prerequisite43dae153; only ancestry changes. The duplicate final-head run was pending at the Human-authorized merge9fc1f8d8; its results are not claimed passing.

avatarSvg(recipe, options)
  ordinary portraits → existing pixelAvatarSvg rendering
  companion surface
-   pixelFigure(recipe, yaw, options)
+   pixelFigure(recipe, yaw, { ...options, speechMouthVersion: 2 })
    each pre-rendered turn → same explicit version2 option

The public-interface regression compares all five saved/closed/half-open layers unchanged and the full opening changed, verifies ordinary portraits keep default rendering and confirms no recipe mutation. No sound, movement, timing, fallback or expression-tool changes. English/Chinese Release Ledgers record the visible correction.

Formal mouth before/after — actual DSH, Chinese/light/1559 × 865

Before source50b44535 uses formal Avatar0.8.0; after sourcea0096785 uses formal0.10.1 plus the Companion-only version2 option. Same isolated Profile, Bot, saved recipe and full reply text; the ordinary Chat retains earlier history between runs. Each has a newly committed native Bot reply, all four observed mouth states, no historical replay on fresh mount and no console errors.

Before: actual speaking sample	After: actual speaking sample

Before: native dismissal restores expression	After: native dismissal restores expression

Continuous44.633s baseline · Continuous40.600s formal improvement · Allowlisted qualification · Capture attribution and limits.

Recipe, ordinary portrait and body hashes match. All five captured saved/closed/half-open layers match exactly; all five full-open layers change. Native dismissal restores saved mouth in the same document. Screenshots are unfrozen speaking samples, not phase-synchronized full-open frames; use the continuous video to judge transitions and the moving opening. Actual delivery timing and Chat history differ. Videos are silent official MCP screencasts, container-remuxed only. Owned browsers are closed; failed setup attempts are excluded. Human accepted the current effect after reviewing the formal after recording on2026-10-09.

To operate: run the isolated DSH launcher on this branch, pin the illustrated Bot, send a new real DM and inspect full versus half-open speech; close the card to check restoration. No TTS or expression-library changes.

Historical validation

Formal dependency/typecheck153files1093tests, lint/format/ledgers/build, source CI at a0096785 and latest full CI at2ef1106a pass. Independent Standards and Spec reviews find no implementation or media-provenance defects against pinned latest main d55d82f. GitHub renders all four embedded screenshots and exposes both video links. Human accepted the formal recording's full opening, half-open mouth and restored expression on2026-10-09.

Prerequisite integration at 659a1415 resolves the updated #1180 branch conflicts while retaining formal0.10.1, both Unreleased entries and Companion-only rendering. The accepted parent release handler is unchanged. Typecheck, all156Client files/1133tests plus the formal mouth renderer regression, lint/source policy, format/ledgers and production build pass. Independent Standards and Spec integration reviews against fixed latest main bcae2205 each report zero findings. Fresh CI37925588791 passes. Historical mouth recordings retain their source attribution and Human acceptance; the separate fast upward throw was accepted on parent source0eb841dd and preserved unchanged by this integration. Broader #1173 background/stale/fallback qualifications remain open.