Stream responder output progressively in talk UI

[ousamabenyounes]

Summary

• Stream responder LLM stdout into an ephemeral per-trigger partial file under the identity run directory.
• Expose active partial replies through the chat API without persisting them to trajectory history.
• Render partial replies in the talk UI as the existing incoming bubble, keep fast polling while partials are active, and suppress duplicate typing/timeout affordances.
• Treat nonzero LLM exits as reply failures even if stdout contains partial text or NO_REPLY.

Fixes #75

RED / GREEN proof

RED on upstream origin/main with the new responder regression:

FAIL partial reply is visible while llm is still running -- missing /tmp/tmp.6xu3gVi946/id/run/responder-partials/trig-1/content.txt
3 passed, 1 failed

RED on upstream origin/main with the new chat API regression:

FAILED web/tests/test_chat_api.py::test_chat_includes_active_partial_reply
1 failed in 0.98s

GREEN on this branch:

test_responder_partial_reply.sh: 12 passed, 0 failed
web/tests/test_chat_api.py: 18 passed

Validation

full local validation suite baseline on origin/main: pass
full local validation suite final on 05d283b: pass
bash suite: passed 40, failed 0
web pytest: 178 passed, 2 skipped
python diff coverage: 44/44 added lines covered
viewer test: 13 passed
viewer typecheck: pass
viewer build: pass

Browser/API proof against a served production build:

partial_text=partial reply in progress
chat_request_count=2
typing_dot_count=0
timeout_note_count=0

Review notes

Independent review found and this branch covers these edge cases:

• partial stdout from a failed LLM call is never sent as the final reply
• failed NO_REPLY output records reply-failed, not permanent no-reply
• retry partials remain visible after a prior reply failed outcome
• partial files tolerate incomplete UTF-8 writes during polling
• malformed step IDs cannot make partial cleanup delete unrelated run/ files