Close the blockers from the fifth audit

Docs and disclosure. HOW-TO-USE.md's settings table still described the
pre-fix behaviour — and help_commands.rs fetches that file from GitHub
main at runtime, ahead of the embedded copy, so it would have reached
every user's Help dialog the moment this merged. The Config tab named
three settings that need a base-image update; there are four, and the
omitted one (Session recap) is the one that fails *without* the "won't
switch off" symptom the warning teaches. Both now also state the cost
nobody had written down: changing any of these recreates the container,
which commits a layer.

Two stale comments that told a reviewer the code was safe when it was
not. compute_claude_code_settings_fingerprint still claimed the
historical fingerprint is preserved so an upgrade cannot churn every
container — carried over from before the widening, false since the
format string changed. And capabilities/default.json, which is the
reviewed threat model of record, described a "Save to host…" action this
branch deletes.

Security and correctness. update_settings validated env vars and nothing
else, so the *global* default_ssh_key_path — the fallback for every
project without an override — took `/` and read-only bind-mounted the
host, which entrypoint.sh then copies into the home volume. classify_
mount_source ran canonicalize on the raw string, which resolves a
relative path against Triple-C's own cwd, so `.` and `..` were accepted
or refused depending on where the app was launched; the daemon then
refuses the mount and the project can never start. Its test passed only
because its examples did not exist under app/src-tauri.

bind_mount_exclusions still derived a path from every row while
project_path_mounts had learned to skip unmountable ones, so a legacy
row made /workspace/<name> ordinary container content that a migration
would then exclude from staging and destroy. The skip is also logged now
rather than silently dropping a folder.

The terminal's file-in path checked is_dir() but not file type, so a
dropped FIFO blocked forever with no timeout — and it is the only route
in now. The web terminal labelled sessions from a global set at request
time, so two quick opens swapped them; harmless until Shift+Enter became
type-dependent, at which point a mislabelled Claude session submitted a
half-written prompt. Opened now carries the type.

Every ~/.claude.json write goes through one atomic helper. The
awsAuthRefresh branches still truncated in place — the same corruption
the Shift+Enter block was fixed for twenty lines later, and its own
comment said so. Demonstrated: a failed write now leaves the original
byte-identical.

And the registration test I added yesterday could pass while the
property was false: an audit got five real unregistered commands past its
exact-string attribute match, and "exactly once" was in its name but not
its body. Mutation-checked against all six shapes.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GBq2rGum6GX7xXgsas1fDc

HOW-TO-USE.md

@@ -615,18 +615,27 @@ The **Claude Code settings** editor, also at the bottom of the Config tab, confi
 
 | Setting | What It Does |
 |---------|-------------|
-| **TUI Mode** | Set to **Fullscreen** for flicker-free alt-screen rendering (uses `CLAUDE_CODE_NO_FLICKER=1`) |
-| **Effort Level** | Controls reasoning depth: **Low** (fast, less thorough), **Medium**, **High** (deep reasoning) |
-| **Focus Mode** | Collapses tool output to one-line summaries, showing only the prompt and final response |
-| **Thinking Summaries** | Shows Claude's thinking process as summaries during responses |
-| **Session Recap** | Provides context when returning to a session after being away |
-| **Auto-Scroll Disabled** | Disables auto-scroll when in fullscreen TUI mode |
+| **TUI Mode** | **Automatic** lets Claude Code choose; **Classic** pins the main-screen renderer; **Fullscreen** pins the flicker-free alt-screen one |
+| **Effort Level** | Reasoning depth: **Low**, **Medium**, **High**, **Extra high** |
+| **Focus Mode** | Summarises tool *calls* to one line each, showing the last prompt and the final response. **Needs the fullscreen renderer** — set TUI Mode to Fullscreen or this does nothing |
+| **Thinking Summaries** | Shows Claude's thinking as summaries rather than a collapsed stub |
+| **Session Recap** | A one-line recap when you return to the terminal after a few minutes away. **On by default** — the switch is how you turn it off |
+| **Auto-Scroll** | Follows new output to the bottom in fullscreen rendering. On by default |
 | **Env Scrub** | Strips credentials from subprocess environments for security |
-| **Prompt Caching (1h)** | Enables 1-hour prompt cache TTL instead of the default 5 minutes |
+| **Prompt Caching (1h)** | Requests a 1-hour prompt cache TTL instead of the default 5 minutes |
 
-Per-project settings override global defaults set in Settings. If all settings are at their defaults, no configuration is injected.
+Each switch has three states on a project: **Global** (follow Settings), **On**, and **Off**. Off is a
+real choice — it overrides a global On, which a project could not previously do.
 
-> These settings map to Claude Code environment variables and `~/.claude/settings.json` entries. Changes require stopping and restarting the container to take effect.
+> These map to Claude Code environment variables and `~/.claude/settings.json` keys, and are applied
+> when the container starts. Changing one stops and recreates the container.
+>
+> **Two caveats on an existing project.** Changing any of these recreates the container, and a
+> recreation commits a new image layer — so flipping switches repeatedly costs disk. And
+> **TUI Mode, Effort Level, Focus Mode and Session Recap cannot be returned to Global** until the
+> project's base image is updated: those four are cleared by *removing* a key, and an older image's
+> startup script ignores the instruction to remove it. Update the base image from the project's
+> Overview tab first. The other switches work on any image.
 
 ### MCP Servers