refactor(web): replace RunningBorder SVG arc with CSS body-shimmer

[lsm]

Replaces the heavy SVG "chasing glow" running indicator with a lightweight pure-CSS shimmer sweep across the card body (the "option 4" mockup the user approved). No border arc, no SVG, no ResizeObserver.

Changes

Visual — while a block is running, a faint white left→right gradient sweeps across the card surface (~1.8s loop). Motion-only signal per the user's decision (no static accent pairing).

New CSS (packages/web/src/styles.css):

• @keyframes runningShimmer + .running-shimmer overlay — position:absolute; inset:0; pointer-events:none, white gradient, background-size:200%, left→right sweep.
• Loop seam is invisible: with background-size:200%, animating background-position 130% → -30% sweeps the band fully off the left edge, across the card, and fully off the right edge.
• prefers-reduced-motion hides it entirely.

Blocks (keep existing isRunning gating intact, only swap the visual):

• ThinkingBlock, SubagentBlock, ToolResultCard — drop the <RunningBorder> wrapper; render the overlay as a card child while running and add relative to the card when running (overflow-hidden clips it to the rounded surface).

Cleanup:

• Delete packages/web/src/components/sdk/RunningBorder.tsx and its 3 imports.
• Update doc/comment references that mentioned the SVG arc / RunningBorder to describe the shimmer (SDKMessageRenderer, SDKAssistantMessage, SubagentBlock, ThinkingBlock, ToolResultCard, tool-types, styles.css).

Tests:

• No existing tests asserted on the arc markup (verified — all svg queries were for icons).
• Added ThinkingBlock test: .running-shimmer is present while running, absent when not.

Verification

• `bun run check` — lint + typecheck + knip + guard scripts all clean (no dangling imports / dead exports).
• `vitest run` on ThinkingBlock / SubagentBlock / ToolResultCard — 180 passed.

[lsm]

🤖 Review by glm-5.1 (GLM)

Model: glm-5.1 | Client: NeoKai | Provider: GLM

Recommendation: REQUEST_CHANGES — 1 P2 finding (one missed stale comment). The refactor itself is correct and well-scoped; all verification passed.

Verified ✅

• bun run check (root: oxlint + tsc --build --noEmit + knip + session-guards + space-task-handler-tests + db-schema-parity + test-quality) — clean, exit 0.
• Block tests green: ThinkingBlock (incl. 2 new shimmer tests), SubagentBlock, SDKAssistantMessage — 130 passed.
• No remaining RunningBorder references anywhere in the repo; RunningBorder.tsx fully deleted; no test imported or asserted on it.
• Shimmer loop math verified: with background-size:200%, the band center sits at W·(1 − P/100); at P=130% → −0.30W (fully off-left) and at P=−30% → +1.30W (fully off-right). Off-screen clearance (0.30W) exceeds the band half-width (0.10W), so the infinite-loop seam is invisible. Sweep is left→right, 1.8s linear. Correct.
• prefers-reduced-motion fully disables it (animation:none + display:none). pointer-events:none + aria-hidden="true" on all three usages.
• Containment: all three card roots are always overflow-hidden and gain relative exactly when isRunning.
• ConfirmModal (ToolResultCard) renders via Modal → <Portal into="body"> + fixed inset-0 z-50, so the in-card shimmer cannot occlude it.
• isRunning plumbing is byte-for-byte identical to before (doc-comment changes only); parent containers don't depend on the old wrapper structure.

Finding

P2 — packages/web/src/components/sdk/SDKAssistantMessage.tsx:261-264 (stale comment, in task scope)

The inline comment still describes the deleted mechanism verbatim:

// When isRunning, ALL blocks in this message receive the animated arc so every block type is visible for debugging/verification. Each component applies the arc via a wrapper div (not directly on its overflow:hidden root) so the inset:-2px extension isn't clipped.

This is now wrong on three counts: there is no "animated arc" (it's a body shimmer), no "wrapper div" (the shimmer is a direct child of the overflow-hidden root), and no "inset:-2px extension". The task explicitly scoped updating arc/RunningBorder doc references in this file — lines 62-66 and 359-366 were updated, but this paragraph was missed. Suggest rewriting it to describe the in-card .running-shimmer overlay (contained by the card's overflow-hidden). Doc-only; no behavioral impact.

Non-blocking notes

• New @keyframes runningShimmer rather than reusing the existing shimmer (skeleton) keyframe: the existing one is px-locked to dark skeleton colors and isn't reusable for a white percentage sweep, so a separate keyframe is the right call.
• Task text said the shimmer should sit "below interactive children"; it's rendered as the last child (paints above content) but pointer-events:none fully preserves interaction, satisfying the intent. No change needed.

[lsm]

🤖 Review by glm-5.1 (GLM)

Model: glm-5.1 | Client: NeoKai | Provider: GLM

Recommendation: APPROVE — zero findings. The P2 doc-comment from the prior round is fixed; all checks and tests pass.

Round-2 verification (fresh, not relying on prior conclusions):

• Fix commit 8bb46f3fe is surgical: only SDKAssistantMessage.tsx (3+/3−). The stale inline comment now correctly describes the in-card .running-shimmer overlay contained by the card's overflow-hidden.
• Independent residual-vocabulary grep across packages/web/src (animated arc / border-light / animateMotion / mpath / radial-gradient mask / inset:-2px / wrapper-div arc / RunningBorder) → only false positives (router *FromPath helpers, unrelated "wrapper div" comments). No real leftovers.
• bun run check (root: oxlint + tsc --build --noEmit + knip + session-guards + space-task-handler-tests + db-schema-parity + test-quality) — exit 0.
• ThinkingBlock / SubagentBlock / SDKAssistantMessage tests — 130 passed.
• PR OPEN, MERGEABLE, head == fix commit (8bb46f3fe), zero unresolved review threads.

The refactor itself (verified in round 1): correct CSS loop-seam math, prefers-reduced-motion, pointer-events:none + aria-hidden, overflow-hidden/relative containment on all three blocks, and ConfirmModal portal isolation. No blocking issues.

(Self-review fallback: GitHub rejects self-APPROVE, so posted as COMMENT with an explicit APPROVE recommendation.)