Sync the skill with the shipped CLI (v0.0.27) - #4

Swept the rest of the machine-readable-output claims. Walked the whole command tree of the installed binary and cross-checked every option against the v0.0.27 tag. The real map:

Two more inventions found and fixed:

csv does not exist. cli-commands.md offered --format FORMAT Output format (table, json, csv) for both ls and ps. lium ls --format csv exits with a usage error — the choice is [table|json].

The ls and ps option lists were largely fictional. ls was documented with --region, --min-memory, --max-price and a positional GPU_TYPE; it has none of them — the real filters are --gpu, --count, --min-cuda, --lat/--lon/--max-distance, --sort, --limit. ps was documented with -a/--all and --sort FIELD; it has neither, only an optional pod id and --format. Both blocks now match the CLI, and the example lium ls --max-price 2.5 is gone.

SKILL.md now carries the table above instead of a one-line summary, and llms-full.txt is regenerated.

Deliberately not documented yet: exec --json landed in DAH-2556 on main but is not in v0.0.27, so the skill would be promising a flag users do not have.

Out of scope for this PR, worth its own pass: the inventions are not limited to output formats — lium ssh is documented with --command, --port and --key, and it takes none of them. A full sweep of cli-commands.md against the CLI is a separate change.

Full sync pass. Method: checked out the v0.0.27 tag, loaded the Click tree programmatically and dumped every command with its real arguments and options — 93 commands — then checked the skill against that dump rather than against memory.

Commands whose documented options did not exist

rm --in / --at were described as name and executor filters; they schedule the removal after a duration or at a time.

Wrong in the other direction

lium init was marked "NOT suitable for agent/scripted use — has no non-interactive flags" while SKILL.md documents the headless flow two sections earlier. It has --no-browser and --session; the reference now says so.

Commands that were missing entirely

balance, topup currencies, topup create, update, ssh-keys — all agent-relevant, none documented. Added.

Exit codes

The table promising codes 1–6 describes main, not the shipped CLI. In v0.0.27 there is no exit-code contract at all: lium ps <unknown-id> exits 0. Verified by running it against the tagged code. The section now states that plainly and keeps the table as what lands in the next release.

Smaller fixes

• lium ls H100 — there is no positional GPU argument, only --gpu;
• --sort default is download, not price_gpu;
• the --json row of the output table named fund crypto currencies / fund crypto invoice, which do not exist — the real ones are fund, topup currencies, topup create;
• sdk-reference.md: lium.exec(pod, "cmd") — command is keyword-only, so the positional form raises TypeError.

Verification

Wrote a checker that extracts every lium … example from the skill and validates each command path and flag against the dump. All examples pass; the only two hits left are prose lines that begin with the word "lium". Every SDK method named in sdk-reference.md exists in v0.0.27, and the call signatures now match.

The checker is a throwaway script, not part of this PR. Worth landing as a CI check next to check-llms-full so the reference cannot drift again — that is its own change.

Superseded by #5, which re-verifies the same claims against v0.0.29 and carries the sdk-reference.md fix from here. Closing.