DAH-2592: describe the real lium CLI in the agent skill

[arhangel66]

DAH-2592. Rewrites the agent-facing CLI reference against the shipped binary
(lium v0.0.29). Documentation only — no CLI code.

Why

lium/references/cli-commands.md did not merely lag the CLI, it invented parts of
it: ls --region/--min-memory/--max-price and a positional GPU_TYPE, ps -a/--sort,
ssh --command/--port/--key, exec --timeout/--output, scp -r/-p,
rsync --delete/--exclude/--dry-run, themes monokai/solarized/dracula/nord where
the CLI takes dark|light. It also denied two capabilities that exist — "lium rm
has no -y, use echo y |" and "lium ps does not support --format json". A
denial is worse than silence: it stops an agent using a command that works.

Seven commands (balance, topup, update, ssh-keys, mine, provider,
gpu-splitting) were missing entirely, and real flags went undocumented —
up --no-ssh/--dockerfile/--ssh-name/--volume-encryption, fund --alpha/-k/--json,
exec -s/--script,-e/--env,--json, ls --count/--min-cuda/--sort/--limit,
ps [POD_ID], scp -d/--download, config path/reset/unset, bk restore-logs.

How it was verified

lium <cmd> --help was captured for all 27 top-level commands and all 62
sub-commands, and every section is diffed against its dump in both directions:
each real flag must appear in the doc, and each flag token in the doc must exist
somewhere in the real help. Sentences that deliberately state a flag does not
exist are exempt — they are what stops an agent re-inventing it.

Behavioural claims that --help cannot answer were read from the source instead.
That is where the two sharpest fixes came from:

• lium ls exits 0 when it fails. GetExecutorsAction swallows every
  exception and the command prints it and returns, so lium ls >/dev/null && echo OK
  reports OK with a revoked API key. Both files used exactly that as their auth
  check; they now use lium balance. Only exec, rm and up are reliable — the
  rest is DAH-2593.
• ps and logs do not accept an index. They compare the argument literally
  against pod ids and names, so lium ps 1 reports "not found". Pod Targeting is now
  a per-command table rather than one "all of them accept" claim.

Also corrected: --sort price_gpu was labelled the default (it is download);
lium up needs --no-ssh or it ends in an interactive SSH session and stalls an
agent; exit codes 3 and 6 are defined but unreachable, so the table publishes
0/1/2/4/5 plus an explicit warning about exit 0; lium mine is provider-side, not
renter-side; LIUM_SSH_KEY does not exist.

llms-full.txt is regenerated in the same change, as check-llms-full.yml requires.

Parallel PRs cover the same drift in lium-docs and in the lium-toolkit:cli
plugin skill.

Supersedes #4

#4 synced the same reference against v0.0.27 from a base that predates #3. Every
claim it fixed is covered here and re-verified against v0.0.29: ps does take
--format json, target pods by name rather than by index, theme accepts only
dark|light, LIUM_SSH_KEY does not exist, and the machine-readable-output map is
published as a table.

The one thing #4 carried that this branch did not is its sdk-reference.md
one-liner, now included. Lium.exec is exec(self, pod, *, command, ...), so the
documented lium.exec(pod, "python /root/train.py") raises TypeError.

Checking the rest of that file for the same defect found it in fifteen more
signatures — scp, upload, download, rsync, switch_template, restore,
backup_now, install_jupyter, schedule_termination and the rest all take
keyword-only arguments the table showed as positional. Two parameters did not exist
at all (exec(..., timeout=), backup_create(..., name=)) and restore's
target_path is really restore_path. Signatures now use Python's own *
notation, with a line above the tables explaining it.

#4 can be closed.