fix: prevent session poisoning from empty partial-stream-stub assistant turns

[0xprincess]

Problem

A mid-tool-call stream drop that delivers no text produces a partial-stream stub carrying content:'' and tool_calls=None. The conversation loop's finish_reason == 'length' truncation path took the no tool calls branch and appended it to messages as {"role":"assistant","content":""} before the continuation nudge, then persisted it.

Strict providers — Moonshot/Kimi via OpenRouter — reject empty assistant content on the next replay:

HTTP 400: Invalid request: the message at position N with role 'assistant' must not be empty

Because the message is persisted, every subsequent turn re-failed — the session was unrecoverable, and resuming it later hit the same wall.

Root cause chain

1. SSE stream ends with no finish_reason mid-tool-call args → _build_partial_stream_stub (content '', tool_calls=None).
2. conversation_loop length path appends the stub via build_assistant_message → empty assistant turn enters history + state.db.
3. Next request replays history → Moonshot 400 → classified non-retryable → session poisoned.

Fix — three layers, smallest blast radius first

1. conversation_loop (length path): an empty partial-stream stub is no longer appended as an interim assistant message — only the continuation user-message is. Stubs that delivered partial text are still persisted, so continuation stitching (existing #32086 behavior) is unchanged.

2. chat_completion_helpers.build_assistant_message: never serialize a textless assistant turn with content:'' — pad to a single space, the same trick as the existing reasoning_content pad (#15250, #17400). Tool-call turns are exempt (content:'' alongside tool_calls is accepted everywhere).

3. conversation_loop send boundary (durable repair): pad a textless assistant turn's empty content to a single space after all content-mutating passes (surrogate sanitize, whitespace normalization, thinking-only drops), before token estimation. This repairs sessions already poisoned by older builds: the persisted stub rows are rebuilt to '' on every reload (_rows_to_conversation strips whitespace, so a DB-side pad can't survive) — only a send-time pad reaches them.

Verification

• 485 tests pass, 0 fail across the four affected files (test_partial_stream_finish_reason.py, test_run_agent.py, test_chat_completion_helpers.py, test_prompt_caching.py, test_turn_finalizer_interrupt_alternation.py), including new regressions reproducing the exact poisoned history shape.
• Live replay: a real poisoned session's resumed history (loaded via SessionDB, repaired, serialized through the patched loop) replayed against Moonshot via OpenRouter returns HTTP 200 (was HTTP 400 before the fix).
• Two pre-existing tests in test_run_agent.py (test_empty_content, test_unterminated_think_block_stripped) asserted the old content == '' behavior — updated to expect ' ' with a comment explaining why.

Notes

• No behavior change for tool-call turns, non-empty turns, or providers that already accept empty content.
• The send-time pad mutates only the per-request API copy; the stored transcript keeps original bytes.

[teknium1]

Merged via PR #73028 (rebase-merge) with your two commits cherry-picked and authorship preserved — thank you @0xprincess, this was a well-diagnosed fix with excellent test coverage. On top of your work we added two follow-ups needed for current main: the multimodal test adapted to main's assistant-content flattening, and a codex_responses exemption (commentary-phase turns persist content:'' by design). All three layers landed as you designed them. Commits 309f06b / 2a26be6 on main.