fix: time to first token for generations that stream no text

[milanagm]

Fixes LFE-16664

completionStartTime was only set when a message_update carried a non-empty text part, so a generation that never streams prose reported no TTFT at all — a model that calls a tool straight away, or one that reasons before answering. The gate now accepts the first delta of any kind: text, reasoning or tool call, since each is a streamed completion token.

Measured in a real pi project (90 days, 49 generations): 13 of 49 (27 %) carried no TTFT, and every one was a tool-call-only completion. The gate has been in place since the plugin's first commit.

Verification

• Two end-to-end tests, one per trigger, driven by the new mock scenarios [tool-only] (tool call, no text, no reasoning) and [think-only] (reasoning then tool call, no text). Both fail against the previous gate. Each also asserts its own premise — no text part, no reasoning part — so it cannot keep passing while no longer covering the bug.
• 130/130 tests, tsc --noEmit clean. main is 128; the two new tests are the whole delta.

Relation to #39

Supersedes the TTFT half of #39, which extended the same gate to reasoning only and explicitly left tool-call-only generations broken ("There is no token to time there") — i.e. 100 % of the cases actually observed. If this lands, #39 narrows to the inline-<think> split.