docs(falcon-ai,protect): add thirteen captures and correct five claims

[khushalsonawat]

What

Fills the ten <img> slots on the Falcon AI guides that pointed at files which never existed, adds three new ones, and corrects the prose the captures disproved.

chat-with-falcon-ai — 9

File	On screen	Callout
open-full-page.png	Full page, Falcon AI selected in the sidebar, three conversations in the rail, an answer in the center	none, layout shot
open-shortcut-panel.png	Tracing project list with the floating button and its Falcon AI (⌘K) tooltip	Opens Falcon AI in a side panel
empty-state-quick-actions.png	"How can I help?", the five quick-action chips, the skills row	Quick actions prefill the input
ask-question-send.png	Question typed, Send active	Send activates once you type
context-selector.png	Context menu open on all seven options, Auto checked	Auto reads the current page
attach-file.png	+ menu open: Attach files, Connectors	Opens the file picker
tool-card-parameters-result.png	list_scenarios expanded: Parameters JSON, Result, Full output table	masks only
stop-mid-response.png	Three tool cards, streaming dots, red Stop in place of Send	Stop keeps the partial answer
copy-rate-response.png	The copy / thumbs up / thumbs down row under an answer	Copy, or rate the answer

Elsewhere — 4

• manage-conversations/row-menu.png — row menu open on Rename and Delete
• create-skill.png — Customize panel, Skills selected, Create Skill at the bottom
• connect-mcp-server.png — Falcon AI Connectors settings, empty state
• protect/guides/test-a-guardrail.png — Test Guardrails panel with the PII test chip applied

Why

Both Falcon AI pages shipped with <img> tags whose files were never added, so ten broken images were rendering in production. Filling them meant driving the product, and driving the product turned up five claims that don't match what it does.

What cases does this cover

• A tool call that took no arguments has no Parameters section at all, so a reader who opens one and finds only Result isn't looking at a bug
• Parameters is a collapsed disclosure inside the already-expanded card, which is two clicks, not one
• The conversation rail reads back as a list of questions rather than summaries, which is why renaming is worth doing
• The + menu's Connectors entry is a submenu, so "Manage connectors" is not where the old text said it was
• The second chip row above the input is skills, not more quick actions

How

Captured over CDP at 3024×1890, native geometry, annotated per annotation-conventions.md: 286×104 ovals, 3 px white stroke, size-28 Arial, solid equilateral arrowheads, straight arrows only.

Every correction is something I saw rather than inferred. Where I could not verify a behaviour without a native file dialog or a config change on the workspace, the page names the control and stops there rather than claiming what it does.

Masking was needed on one image: the agent_id in the Parameters JSON, the agent UUID in the "Showing 1 of 1" line, and the scenario ID in the table.

Known gaps

Two rows from the same sweep are blocked on data, not on effort:

• choose-connector-tools needs a connector actually connected with tools discovered
• review-guardrail-activity needs a workspace with real guardrail triggers; Ava's four KPI cards all read zero with all three guardrails disabled, and I did not enable one to manufacture data

Recording

n/a — static pages.

Checklist

• All five pages render locally (200)
• All 13 referenced assets resolve (13/13)
• No unreferenced assets added
• Documentation updated