docs: sync needle run --temperature and finetune --generate options

[LeonSGP43]

What

Two documentation sync gaps between the CLI (as of 4ac96b1 / #88 and 25ef4ac / #93) and the docs:

1. llms.txt CLI summary omits needle run --temperature. The entry lists --checkpoint/--query/--tools, but fix: wire --tools and --temperature through needle run, drop dead --no-constrained flag #88 wired --tools and --temperature through needle run. Only --temperature is missing from the summary.

2. doc/finetuning.md does not mention the inline generator options of needle finetune. --generate <n> (synthesize N extra examples via OpenRouter before training), --model <id> (default deepseek/deepseek-v4-flash), and --workers <n> (default 8) are documented in README ("Key options" list) and llms.txt, but the finetuning deep dive — the doc that explains exactly this pipeline, including "when validation loss rises, add data" — never mentions the built-in synthesis path. It only covers the separate generate-data command.

Evidence

• needle/cli.py, run subparser: --temperature, type=float, default=0.0, help="Sampling temperature (0 = greedy)".
• needle/model/run.py generate(): temperature <= 0.0 → argmax (greedy); above 0 → jax.random.categorical(logits / temperature).
• needle/cli.py, finetune subparser: --generate (default=0, "Generate N extra examples via OpenRouter before training (0 = off)"), --model (default="deepseek/deepseek-v4-flash"), --workers (default=8).
• README already covers all three finetune options; doc/finetuning.md and llms.txt were the two files out of sync.

Changes

• llms.txt: add [--temperature <t>] to the needle run entry with a short "defaults to 0 (greedy)" clause, matching the bracketed-optional-flag style of the neighbouring entries.
• doc/finetuning.md: one sentence after the generate-data block describing --generate/--model/--workers, using only facts verifiable from the CLI definitions.

Docs only; no code paths touched.

[LeonSGP43]

Hi! Just following up briefly on this one — it's been a couple of days since submission, so wanted to check it's on the review radar. Both syncs (needle run --temperature in llms.txt, finetune --generate/--model/--workers in doc/finetuning.md) were verified against the current source (ee221ce). Happy to adjust the scope or split it into separate PRs if that's easier to review. Thanks!