[addisdev]

Replaces the last three mermaid diagrams in the docs with hand-placed figures, brings the workloads figure up to the catalog, and re-shoots the Devices screen and adds a Findings screen from the live brain. No mermaid is left under docs/, and CI now keeps it that way.

Before / after

Page	Before	After
explore/index.md	mermaid flowchart TD, in GitHub's/Material's theme	explore-night.png: the walk, the replay, and the verdict loop that switches a check off
explore/how-it-works.md	mermaid sequenceDiagram	explore-step.png: one mission across executor, device, model, judge and collector
explore/internals.md	mermaid flowchart TB with subgraph	explore-code.png: the code map, plus the "pixels outside model.ts" rule
README, workloads/index.md	workloads.png: 28 workloads; README said "the figure predates six of them"	workloads.png: all 34 shipped catalog entries; the aside is gone
dashboard.md	devices.png: a throwaway laptop + simulator, generic glyphs	devices.png: the PR #69 drawings on the real shelf (MacBooks, Galaxy, Roku in front of a TV, a sim)
dashboard.md, explore/findings.md	no Findings picture	findings.png

Every label was checked against the code rather than the old diagrams: two replays (replay_attempts), filed on at least 1 of 2 or a crash log (fileable()), precision under 30% once ten are judged (PRECISION_FLOOR), and pilot/vision as the gateway defaults. The old diagrams had two claims I did not keep: "the model" as the only model (the judge is one too), and an implied "same model never judges", which the code defaults but does not enforce. Figures follow the house rules: 2x PNG on their own charcoal ground, brand.css + symbols.js, nothing under 14 px at 1x.

The dashboard shots

npm run shoot:dash cannot capture either: its throwaway collector has one machine agent (so no shelf to draw) and no explore night (so no findings), and seeding either would be inventing data. Its three images (first-result, alerts, visual) show neither screen, so I did not re-run it.

Instead, a new read-only npm run shoot:live -- --url <brain> opens two pages on the running brain and writes nothing to it:

• Devices filtered to status=online (5 of 20 registered). The full list is 15 offline, which brand.md already records reading as abandoned.
• Findings filtered to app=bug-garden, the fixture app with planted defects. Real GreenFolio findings were left out on purpose; publishing a shipping app's open bugs is your call.

Both filters show in the picture. The banner's four open alerts are real (shelf devices that have not checked in). brand.md says all of this.

CI and config

• docs.yml has a new No mermaid step that fails on a mermaid fence in any docs/**/*.md. Tested against a scratch file containing one.
• mkdocs.yml drops the mermaid custom_fences and keeps plain pymdownx.superfences, which tabs and admonitions need.

Verified

• mkdocs build --strict passes on top of Docs catch up with 0.8.0 and the explore work #88, and all six images resolve in the built site.
• tsc --noEmit in collector/ is clean.
• npm test in collector/ passes (ALL PASS, 1116 ok, with the one usual skip: no runner-android APK built). shoot:live is not part of the suite; I ran it by hand to make the two captures.

Rebased onto develop after #88 merged, so there are no conflicts with it.

🤖 Generated with Claude Code

[addisdev]

Reviewed the rendered figures (explore-night, explore-step, Findings capture): accurate to the code and on-brand, live data only. CI green. Merging.