[santoshkumarradha]

What was wrong

A factory chat with no checkout of its own fell back to its host's workspace, and that was never a project:

• cmd/codeaf/appfactory_setup.go — the desk's own factory host was built with workspace: home.Dir(), so the codeaf home ITSELF, whatever the engine's cwd. This is the path the WorkAF desk takes.
• cmd/codeaf/app_factory.go (newAppFactoryHost), cmd/codeaf/runner_command.go — os.Getwd(); cmd/codeaf/factory_run.go — the launch's workspace.
• cmd/codeaf/factory_agentfield_run.go afRemote.mirror minted every AgentField step's chat with v3MintSession(bucket, r.workspace, r.workspace) and wrote r.workspace as the header's cwd, so every SWE-AF / RunOn step chat worked in ~/.codeaf.
• The same fallback fed the foreman (foremanMaker) and a stage with no worktree (stageMaker.Open).

Every save of those chats ran the seal guard's secret scan over ~/.codeaf — graph.db, factory clones, bin/, credentials.json — into the key vault.

The fix

• AgentField-run jobs work in the job's own empty folder, <floor>/chats/<job> — the folder talkFolder already used as its last resort, so no new layout. Step mirror chats always use it. A RunOn job's or all-AgentField pipeline's manager uses it too (factory.RunsElsewhere), even when this machine knows a checkout. The chat layer has no "no folder" mode (v3MintSession needs a workspace, and an empty one becomes a project bucket keyed on .), so an explicit empty folder is the option the code supports.
• Hosts drop a workspace that is the codeaf home, above it or inside it (factoryHostWorkspace); the desk host now has none. The foreman falls back to <floor>/chats/foreman; a stage with no worktree falls back to its job's folder.
• Guard: factoryFolderRefusal is asked by every factory mint site (talk/manager, stage, foreman, AgentField step). It refuses the codeaf home, any folder above it, and any folder inside it that is not strictly inside the floor's work/, repos/ or chats/. The error is a factory chat cannot work in codeaf's own folder (<dir>). Local jobs' worktrees, clones and recorded checkouts pass unchanged.
• Existing chats that already name ~/.codeaf: internal/cellstore holdsHome now treats the codeaf home and its ancestors the way it treats the person's home. Such a chat seals only its own folder, so the scan never walks the home again, whichever door reopens it. Its tools still run there, and nothing is moved or deleted. It shows one notice of its own: "this folder holds codeaf's own folder, so its files are not saved with this chat · …". A folder inside the home, such as a job worktree, still seals as a project.

Manual

• factory-agentfield.md: new section, "which folder an AgentField job's chat works in".
• sessions-and-rewind.md: new section, "A chat opened in codeaf's own folder", which quotes the new notice.

Tests

• cmd/codeaf/factory_chat_folder_test.go. With cwd = the codeaf home, an AgentField step's chat works in <floor>/chats/<id> (empty), not in cwd, not in a folder holding the home, and the header names it. A RunOn manager uses its own folder even when a checkout is known. Local job: its checkout, worktree and clone pass. The guard refuses the home, its parent, v3, the floor, repos, work and bin. A host started in the home gets no workspace. The foreman falls back to chats/foreman.
• internal/cellstore/codeaf_home_test.go: the codeaf home and its parent are never sealed; a job worktree inside it is.
• make pr-ready BASE=origin/app/v3 passed: the light gate, all of ./cmd/codeaf (8 shards) and all of ./internal/cellstore, every test on its first run. internal/manual and internal/namelaw are green too.

Touches near open PRs

internal/cellstore/engine.go and seat.go only. #2576 changes captureplan.go and guard.go, so the files do not overlap. Nothing under internal/factory/agentfield (#2573) is touched.