fix: avoid gluing reasoning sentences across deltas

[iqdoctor]

Why

Some reasoning streams arrive in sentence-sized deltas.

Example:

• approach.
• Specifying session retrieval format

fast-agent currently concatenates them raw, which yields:

• approach.Specifying ...

That is not a formatting preference issue. It corrupts human-readable reasoning in the console.

Change

Normalize reasoning deltas at the provider boundary.

Rules are intentionally small:

• insert one space when a new reasoning delta starts a new sentence/token boundary
• do not insert spaces for call-like fragments such as voice.fetch(...)

This keeps the fix local to reasoning stream assembly and avoids adding heuristics in the UI layer.

Why this shape

Codex-style stream handling is simpler when text is fixed close to the event source.
UI should render, not repair token boundaries.

Validation

• added focused tests for reasoning delta normalization
• uv run scripts/lint.py
• uv run scripts/typecheck.py