docs(simulation): capture Fix My Agent and correct four claims

[khushalsonawat]

What

The Fix My Agent page had no images at all. Its tracker row was filed as a video, so it never entered the screenshot sweep and the gap went unnoticed until now.

Adds one capture that carries the whole page in a single frame: the panel open beside the tabs, the run Summary, the suggestion count, Optimize My Agent, the Agent Level / Branch Level toggle, and two branch-level issues with their priority chip, recommendation, branch category and affected calls. One callout, on the toggle. Both execution identifiers masked.

Driving the panel to get that shot turned up four wrong claims and one missing behaviour.

Why

One image is the right number here. Every state this page describes lives in the same panel, so a second shot would have been the same screenshot with a different row highlighted, which is the pattern the reader-analyst passes on #814 called out as training people to stop looking at images.

What cases does this cover

• There is no domain level. The page said you switch between issues at the agent level and the domain level. The toggle is Agent Level and Branch Level, and a branch-level entry names its path on a Branch Category line the page never mentioned
• The counts are nested, not two. Suggestions (n) heads the group, and the per-tab count says how many of that total the current tab is showing. Verified they sum: Agent Level 1 + Branch Level 4 against a heading of 5, on a run whose other execution showed 7
• Infra based suggestions is about runtime, not setup. The three entries are Reduce Perceived Latency, Mitigate Latency Spikes and Prevent Conversation Looping. The page called it "how the agent is set up". Its entries also carry the same priority, recommendation and calls as the prompt-based ones, and the group opens with a Human Comparison Summary
• The collapse chevron is at the panel's left edge. Without that, clicking Fix My Agent reads as doing nothing. It cost me two wrong clicks getting this capture
• An analysis is kept per execution, stamped Last updated at beside the count, so reopening shows the stored result rather than starting again. That is what makes the empty state first-time-only

How

Captured over CDP at 3024×1890 on a completed 100-chat run, annotated per annotation-conventions.md. The oval was placed from an empty_boxes scan rather than by eye, and sits on canvas measuring flat and dark.

Claims were checked rather than assumed:

• selecting an issue does narrow the grid. The header switches from All Chats (100) to (6) Chats selected, matching that suggestion's Calls Affected (6)
• the panel does survive a tab switch, verified by moving to Analytics with it open, which is what the caption rests on
• the two suggestion counts are the same set, verified by arithmetic across two executions

Masking: both execution identifiers, in the breadcrumb and the heading.

Known gaps

• The empty-state string remains unverified. The page quotes "There are no suggestions yet, click the refresh button to get suggestions." Every execution reachable on this run has already been analysed, so that state was not reachable. The string is left as the page's existing claim rather than presented as checked, and the new caching paragraph explains why it is rare
• A third execution would not open the panel across four attempts. I could not separate a product fault from a flaky synthetic click, so nothing is claimed about it
• I did not click refresh: it re-runs the analysis, which costs credits and takes several minutes

Recording

n/a — static page.

Checklist

• Page renders locally (200)
• Referenced asset resolves
• Oval verified clear of UI against the source
• Identifiers masked
• Documentation updated