fix(process_registry): poll() must not mark completion as consumed (#10156)

[someaka]

Problem

poll() in process_registry.py calls _completion_consumed.add(session_id) when the queried process has already exited. This marks the completion as consumed, which causes the gateway watcher (and notify_on_complete drain loops in the CLI and TUI) to suppress the completion notification — they check is_completion_consumed() and skip it.

poll() is a read-only status query, analogous to checking whether a letter has arrived without opening the envelope. It should never suppress the eventual delivery notification.

Root cause

PR #8228 introduced _completion_consumed tracking to suppress duplicate completion notifications when the agent has already seen the output via wait() or read_log(). The fix correctly applied to wait() and read_log() (both are agent-facing read operations that consume the result), but incorrectly also added the mark to poll().

Fix

Remove the single line self._completion_consumed.add(session_id) from poll(). Only wait() and read_log() mark completions as consumed.

Tests

• test_poll_does_not_mark_completion_consumed — renamed from test_poll_marks_completion_consumed, assertion flipped to assert not
• test_poll_then_wait_marks_consumed — new: poll() followed by wait() — poll() must not consume, wait() must still consume
• test_poll_then_read_log_marks_consumed — new: poll() followed by read_log() — poll() must not consume, read_log() must still consume
• Existing test_wait_marks_completion_consumed and test_log_marks_completion_consumed unchanged

Fixes #10156

[alt-glitch]

Duplicate of #10158 (and #21903). All three fix the same poll() marking _completion_consumed regression from #8228 — making poll() read-only so notify_on_complete watcher is not suppressed. Fixes #10156.

[someaka]

Closing as duplicate — #10158 and #21903 already address the same regression with the same fix (making poll() read-only).