fix: surface two silently-dropped configs

[pedropaf]

Two bugs found while running a 36-image prompt experiment today. They share a failure mode: the config looks accepted, nothing errors, and the run is quietly not what was asked for. That's the worst kind of bug for batch work — you only find out by auditing output counts.

1. seeds: under defaults: was silently discarded

StepDefaults has seed (singular) and no deny_unknown_fields, so serde dropped seeds: without complaint.

defaults:
  seeds: [7, 42, 1234]   # ignored — 1 image per step, not 3

I hit this designing a seed-replication run. It only surfaced because I happened to --dry-run first and noticed count=1. Had I not, I'd have drawn conclusions from a third of the intended data.

Unknown keys are now captured via #[serde(flatten)] and warned at parse time, with a specific hint for seeds since it's step-level only:

! `seeds` is not a workflow default — it is only valid on a step, and was ignored here.
  Move `seeds: [...]` onto each step.
! unknown key `bogus` under `defaults:` — ignored

Warn rather than deny_unknown_fields so existing workflows carrying stray keys keep running. This catches every future typo under defaults:, not just this one.

2. Missing persistent worker was invisible

try_submit_via_socket swallowed the connect error and fell back to one-shot mode. Without a running daemon every job reloads the model from scratch — indefinitely, with no indication. Every run I did today paid this, and I only noticed by reading executor.rs.

Now hints once per process:

i No persistent worker — each job reloads the model.
  Start one with `modl worker start` to keep it in VRAM.

Not auto-started: that's a bigger decision about VRAM ownership and belongs in its own change.

3. Docs: --skip-existing doesn't exist

skills/modl/SKILL.md documented a flag that was never implemented. Skipping already-generated sub-jobs is the default; --force opts out. (The same error is in the org-level CLAUDE.md, which isn't in this repo — fixed separately.)

Testing

• 232 tests pass (231 + 1 new regression test pinning the inert-seeds behaviour)
• Both warnings verified firing against a real workflow
• cargo fmt + clippy clean via pre-commit

🤖 Generated with Claude Code