fix(desktop/transcript): accurate insight-turn jumps — auto behavior + item-height estimate

[agentfleets]

What

Fixes a director-reported bug in the agent transcript's INSIGHT view: clicking different turns (or error rows) jumped to the same — wrong — position instead of the turn's location.

Two surgical changes, both in desktop/src/surfaces/AgentTranscript.tsx:

1. scrollToSeq now calls scrollToIndex with behavior: 'auto' instead of 'smooth'.
2. The insight Virtuoso gets defaultItemHeight={120} (parity with the live list).

Root cause

Verified against the react-virtuoso 4.18.10 sources (node_modules/react-virtuoso/dist/index.mjs, scrollToIndex system ~:1100-1160):

• For an off-screen (unmeasured) target, Virtuoso computes the target offset from estimated row heights — a running average of the rows measured so far. Transcript cards vary wildly in height, and the insight list had no defaultItemHeight, so early-measured rows skewed the average. Estimates for distant indices overshot, and the browser clamped the scroll at the same max offset (bottom) — hence different turns landing at the same position.
• Virtuoso's smooth path only retries the landing when smoothScrollTargetReached fires, and a hard setTimeout(..., 1200) cleanup cuts that retry chain mid-flight, leaving the list wherever the last estimated fling stopped.
• The 'auto' path instead re-fires scrollToIndex on every listRefresh (debounced 150ms) until the list stops changing — it converges to the true offset as rows get measured.

The 1.4s ev-flash that already follows a jump gives the user the orientation the smooth animation was providing, so nothing is lost visually. The added defaultItemHeight={120} improves the estimate the auto-retry converges from.

Side effect (intentional improvement, not scope creep): the live-mode lens stepper uses the same scrollToSeq, so its long-distance jumps are now accurate too.

How verified

• Root cause verified against react-virtuoso 4.18.10 sources (no device available — CI is the gate).
• npm ci && npm run build (desktop): clean (tsc + vite).
• bash scripts/lint-desktop-tokens.sh: clean (66 off-token, all <= baseline, no phantom tokens).
• git diff origin/main --stat: 1 file changed, 15 insertions(+), 1 deletion(-) — desktop/src/surfaces/AgentTranscript.tsx only.
• No new user-visible strings (i18n untouched); no CSS, manifest, or version changes.

[physercoe]

Reviewed — approve. Focused on what CI can't catch (correctness/regressions).

• Diff matches the description; only two hunks in AgentTranscript.tsx (behavior smooth→auto; defaultItemHeight={120} on the insight Virtuoso).
• scrollToSeq is shared by both the live lens stepper and insight jumps; auto fixes the same estimated-height overshoot for both — consistent, no live-path regression.
• defaultItemHeight={120} matches the live list exactly, so the two instances are now consistent; it only seeds the pre-measurement estimate.
• The 1.4s ev-flash the PR cites for orientation genuinely exists and is applied on jump, so nothing visual is lost by dropping the smooth animation.
• No index/off-by-one/behavior-string bugs; index < 0 guarded.

Non-blocking (pre-existing, not introduced here): a single virtuosoRef is shared by both Virtuoso instances — safe only because exactly one renders per mode.