fix: stabilization pass — fail-closed auth, ai flag validation, lint/config/security hardening, regression tests

17-point code-level audit executed via Explorer->Architect->Builder->Tester->Reviewer;
Reviewer accepted (APPROVE_WITH_NOTES; 3 block-list items resolved):

- security: telegram sender-owner AND-gate + TELEGRAM_OWNER_ID, matrix
  MATRIX_ROOM_ID fail-closed, gpg --passphrase-fd 3 (no argv secret),
  /dev/tcp positional-arg form (checkport/smb-client/share-lib/NET_PROBE),
  eval deny-by-default + --no-command-execution carried by both chat bridges,
  tty-gated --trust; config/{telegram,matrix}.env reference templates
- ai: all ExecStart flags validated against installed llama.cpp
  (requested->error, default->omit+warn, CONFIG_REQUESTED_FLAGS); single-file
  hf download failure rc=1 + no .hf-meta; LLAMACPP_HOST coherent;
  POS_SUBCMDS + metadata gaps closed
- tooling: lint-conventions Bash-native rewrite (~24-30x faster, rules and
  output byte-identical, :num restored); pos system uninstall covers all 12
  libs + scale-tail + flags dir + systemd user units (|| true) + plugin
  markers; anchored .bash_completion/.bashrc removal replaces sed -i '/pos/d'
- config: canonical load_env_file in lib/config-ui.sh (CRLF strip, env-wins,
  XDG, LOADED_ENV_KEYS); 9 tools migrated; entertainment-lib collapsed to
  wrappers; docker-compose deliberately unmigrated (source semantics)
- tests: first committed regression suite — tests/run-tests.sh zero-dep
  runner + make test; 12 files / 179 checks / 0 skip / ~52s; hard skip
  contract; systemd-analyze verify on generated unit PASS

Verified: make gen idempotent; make check green; make lint 0 FAIL, 0 WARN;
make test green; bash -n clean; git diff --check clean. Audit deliverables +
agent reports + AGENT_TODO Done entry included.

DOC/howto/ai.md

@@ -25,6 +25,9 @@ attaches the latest pos command output or captured output (tail, max 4096 chars)
 to the question and notes on stderr which source was attached, its age, and a
 staleness warning once it is older than an hour (`ask` only; stdout stays pure
 answer). Use `capture` to save output from any command for `--last`.
+`--trust` and `--no-command-execution` control how (if at all) agent-detected
+command blocks run — see [Command execution posture](#command-execution-posture)
+below.
 
 Backward compatibility: `pos ai gemini`, `pos ai openrouter`, and
 `pos ai llamacpp` still work as shorthand for `pos ai --provider gemini`,
@@ -245,6 +248,31 @@ truncated). To disable: `unset __POS_CAPTURE_ACTIVE`.
 - On a non-2xx response the API's `error.message` is shown and the exit code is
   non-zero — so scripts can rely on `ask` failing loudly.
 
+## Command execution posture
+
+When a model's answer contains a ```sh/shell fenced code block, `ask`/`chat`
+offer to **run** it as a shell command on your machine. Because that code is
+untrusted, AI-generated external authority, the default is to **deny**:
+
+- On an interactive terminal, `ask`/`chat` print the detected command and prompt
+  `Run this command? [y/N]` — you must type **`y`** (or `Y`) to execute it.
+  Enter or any other key **declines**: the command is added to your shell
+  history (press ↑ to recall and edit it) but never run.
+- Without an interactive controlling tty (pipes, scripts, cron, the Telegram /
+  Matrix bridges) commands are **never** executed — the code block is neither
+  printed nor run.
+
+`--trust` auto-executes detected commands **without** the confirmation prompt —
+but only on an interactive terminal. It is meant for trusted alias wrappers;
+do not pass it unless you fully trust the agent's output. It has no effect in a
+non-tty/chat-bridge context, which never executes commands anyway.
+
+`--no-command-execution` disables execution entirely: the detected command is
+neither printed nor run and no confirmation prompt appears. It is the
+structural guard used by the chat bridges so that a future refactor cannot
+accidentally auto-execute model output. When both `--trust` and
+`--no-command-execution` are given, the last one on the command line wins.
+
 ## Troubleshooting
 
 - `ask` errors "No Gemini API key — run 'pos config ai'" → the key isn't set