Merge pull request #3197 from bobleer/bob/bitblob-gentle-gaze

fix(pets): soften BitBlob directional gaze

Author: bobleer

src/web-ui/public/agent-companion-pets/bitblob/README.md

@@ -15,7 +15,7 @@ A smooth, rounded lavender companion with dark oval eyes and a soft ball antenna
 
 Created with the built-in image generation tool and the `hatch-pet` v2 workflow on 2026-09-21. The canonical prompt specifies a compact lavender blob, dark purple inset oval eyes, a short soft antenna with purple ball, optional small attached hands, smooth soft shading, no mouth or props, no text, shadows or detached effects. The supplied sheet serves only as a visual reference.
 
-Each state was generated as its own grounded strip: quiet breathing/blinking; rightward squash/stretch gliding; connected-hand wave; squash/rise/peak/fall/landing; disappointed antenna droop; expectant hands-together asking pose; focused stationary task processing; calm thoughtful review. Leftward gliding is the sole derived visual row, produced by an approved framewise mirror of the symmetric rightward sequence without reversing timing. The four-cardinal anchors and each eight-pose look row were generated separately. Look directions keep the lower base planted while the upper face and inset eyes turn/stretch toward the target, with gentle antenna follow-through.
+Each state was generated as its own grounded strip: quiet breathing/blinking; rightward squash/stretch gliding; connected-hand wave; squash/rise/peak/fall/landing; disappointed antenna droop; expectant hands-together asking pose; focused stationary task processing; calm thoughtful review. Leftward gliding is the sole derived visual row, produced by an approved framewise mirror of the symmetric rightward sequence without reversing timing. The four-cardinal anchors and each eight-pose look row were generated separately. Look directions keep the lower base planted and the face nearly frontal. Both inset oval eyes remain near the middle of the face, with restrained directional offsets and gentle antenna follow-through.
 
 Deterministic extraction, registration, composition and edge cleanup use the skill's scripts. Jumping uses the supported shared-viewport extraction to preserve the generated vertical arc; other standard rows use connected-component extraction.
 
@@ -30,13 +30,33 @@ is aligned to x=95.5 and the baseline to y=194 in each 192×208 cell. Jump frame
 use offsets of 0, 4, 16, 6, 0 pixels from that baseline.
 
 This corrects the former half-sized jump, oversized working/waiting frames and
-per-frame size drift during review. The original poses, frame order, v2 layout
+per-frame size drift during review. The standard action poses, frame order, v2 layout
 and transparency remain intact. BitBlob plays only the atlas animation in the
 shared renderer; additional CSS breathing, hover squash and work/drag transforms
 are disabled for this built-in pet so they do not compound its authored motion.
 
+## Gentle gaze
+
+The 16 look directions were regenerated as two coherent eight-pose strips on
+2026-09-22, grounded in the existing neutral pose and a new set of restrained
+cardinal anchors. Both eyes remain visible on the central front face; lateral
+looks no longer push them toward the silhouette or turn the body into a profile.
+The downward-to-left-to-upward eye line rises gradually through the second row.
+
+Only rows 9–10 are replaced. Decoded pixels in rows 0–8, including the neutral
+slot, are identical to the preceding size-normalized atlas. The new look cells
+use the same silhouette-area normalization and baseline; their alpha-weighted
+areas vary by less than 2% across the 16 directions.
+
 ## Verification
 
-Validation covers the v2 geometry, populated and unused cells, alpha/chroma cleanup, frame clipping and component extraction. Independent visual review covers all nine animations, every look direction and continuity around the row boundaries. Three isolated reviewers classify the randomized horizontal/vertical direction sheet; cardinal ambiguity blocks packaging. Runtime/transport integration checks belong to the companion feature tests rather than the image pipeline.
+The packaged v2 atlas passes geometry, populated/unused cell, transparency and
+chroma checks. Registration also checks final-cell edges and verifies that all
+standard rows are pixel-identical. Direction continuity measurements and an
+independent visual review cover all 16 poses. Three isolated reviewers classify
+the randomized horizontal/vertical direction pairs. Very small intermediate
+vertical cues are intentionally subtle; cardinal directions remain readable.
 
-The completed run passed v2 atlas validation and a single alpha-preserving edge cleanup, with no atlas errors or warnings. All four cardinal directions passed three-reviewer blind consensus. The subtle horizontal component at 22.5° remains a reviewed warning; labeled review confirms the intended quadrant. The closing upper-left-to-up step and two continuity metrics also remain reviewed visual warnings, without a wrong quadrant, clipping or visible positional snap.
+This is a static artwork replacement with no renderer, persisted data or
+transport changes. Remote workspace, remote control, Peer Device Mode and
+Detached Dispatch were not exercised for this asset-only change.