fix(kimi): stop sending temperature, which Kimi pins per model

[klaus993]

Third and hopefully last link in the chain from #14 -> #15.

What happened

#15 set thinking: 'disabled' on kimi-k2.6. The next /kimi run came back with a 400, posted as a proper error comment (the #15 hardening working as intended):

Review failed: Kimi API request failed with status 400: invalid temperature: only 0.6 is allowed for this model

Disabling thinking is precisely what moved k2.6 into its non-thinking regime, where the temperature: 1 this workflow has always sent is rejected.

The rule

From the Model Parameter Reference:

Parameter	kimi-k3	kimi-k2.7-code	kimi-k2.6
temperature	Fixed at 1.0	Fixed at 1.0	1.0 thinking / 0.6 non-thinking

kimi-k2.6: fixed at 1.0 in thinking mode and 0.6 in non-thinking mode; other values return an error.

Do not pass temperature explicitly when calling these models.

The fix

Stop sending it. Hardcoding 0.6 would be correct only while the model stays k2.6 and thinking stays off -- it would break the moment either moves, which is exactly the trap that produced this PR. Omitting the field is the only value correct across all three models.

temperature becomes a string input defaulting to '', added to the payload only when non-empty, mirroring thinking:

+ (if $temperature == "" then {} else { temperature: ($temperature | tonumber) } end)

Anyone who needs an explicit value can still pass one as a string.

Compatibility

No caller passes temperature. Across all 12 consumers, the only one that passes any input is lambda_compiler_kit (a custom prompt), so changing the input type from number to string breaks nobody.

Verification

YAML parses; inputs resolve to model=kimi-k2.6, thinking=disabled, temperature=''. Payload construction exercised across the thinking x temperature matrix, asserting on has() rather than eyeballing that both keys are genuinely absent (not null) when their input is '', and present with the right value when set.

Not verified against the live API -- that needs the merge plus a v1 move. v1 currently points at d66ba3f (#15), which 400s on every review.