fix: rewrite streaming path via run()+event_stream_handler

[sjwan01]

Summary

Fixes streaming completeness: run_stream() / run_with_events() previously used agent.run_stream(), whose stop-at-first-output semantics truncated the run when the model emitted preamble text alongside tool calls — the post-tool summary was never generated. The streaming path now uses agent.run() + event_stream_handler (the combination pydantic-ai recommends for "run to completion and stream at the same time"): model requests stream (token events) while the graph advances non-streamingly, so tools always execute and the second request always happens.

Also adds a "Working Principles" section to AGENTS.md (think before coding / simplicity first / surgical changes / goal-driven execution).

Changes

• _run_stream_impl: agent.run_stream + stream_text → agent.run() + event_stream_handler as an asyncio.Task, events forwarded through a queue. Token events = text deltas (PartStartEvent(TextPart).content + PartDeltaEvent(TextPartDelta).content_delta). run_end.output = result.output (complete summary, not delta concatenation). Early consumer abort cancels the run task cleanly.
• Thinking warning removed (001's): based on wrong causality (thinking was not the trigger) and obsolete (the defect is fixed). Removed from code, README Known issues, docstrings, tests.
• Version: 0.2.0 → 0.2.1 (patch — bug fix).
• Tests: deterministic FunctionModel tests (preamble + tool calls → complete output, event ordering), early-abort-cancels-run-task, tool-event completeness, run_with_events() completeness. make check: 156 passed, 7 skipped.
• AGENTS.md: new Working Principles section.

Verification

• Real-API end-to-end (deepseek-v4-flash): run_stream output complete 4/4; run_with_events event stream complete + ordered (token/tool/run_end).
• Early-abort cancel verified on pydantic-ai 2.22 and 2.25 (clean cancel + generator closed).
• make check green (ruff / mypy strict / 156 passed, 7 skipped).

Behavior notes (0.2.1)

• TOKEN_STREAM granularity: one event per text delta (preamble included); run_end.output is the complete final summary.
• The thinking warning is gone (it was based on an incorrect causality and the defect is fixed).