fix(server): report a failing end-of-stream terminator instead of swallowing it

[cdeust]

Summary

Fixes #135 — a separate defect from #134, and one that survives it.
graph_build_run.py's finally block sends the single end-of-stream
close() call inside except Exception: pass. Any failure there is
completely invisible: no log line, no counter. That silence is exactly why
#134's AttributeError on this same call (close() didn't exist after
45d4a80) survived a release unnoticed — #134 stops that failure, not the
next one this handler can still hide.

The existing comment's reasoning is half right: a failure here must not
mask the build outcome being unwound, so it must not be re-raised. But
not-masking and not-reporting are different things.

What changed

except Exception: pass → the exception's type and message are printed to
stderr before the lock releases. Still never re-raised.

Tests (tests/test_graph_build_run_terminator.py)

• run_build's first statement is forced to fail (no store/DB dependency
  needed to reach finally).
• close() raising is reported on stderr with the exception named, without
  re-raising and without leaking the build lock. Verified failing against
  the pre-fix except Exception: pass handler.
• The nominal path (close() succeeds) stays quiet — asserted, not assumed.

Verification

• uv run pytest tests/ -k "graph_build or graph_event or activity_stream or http_live" — 74 passed, 2 skipped.
• uv run ruff check / ruff format --check — clean.

Co-Authored-By: Claude noreply@anthropic.com

[cdeust]

ZETETIC-REVIEW: APPROVE

This is the defect that let the other one reach a release, and it is correctly treated as separate rather than folded into it. #136 makes close() resolve again, which stops this failure and not the next one: the handler was written so that any failure to terminate the stream is invisible — no log line, no counter, nothing.

The original comment's reasoning was half right, and the fix keeps the half that was: a failure here must not mask the build outcome being unwound. But not-masking and not-reporting are different things. The terminator now prints the exception type and message to stderr without re-raising, so the unwind is preserved and the silence is not.

The consequence of the silence was real, not theoretical. Subscribers never received done, so they never disconnected and sat on an open SSE connection waiting for a terminator that was never sent — after a build that had spent fifteen minutes getting there. And at the level of the incident: the loud error was at line 155 and the second one at the bottom made no sound at all, which is how a build that could not start survived a release.

Both directions are tested, and the failing one was proven against the pre-fix code. One test forces run_build's first statement to fail so finally is reached without a database dependency, and asserts the failure is now reported and the lock still releases — verified failing against the original except Exception: pass by stashing back to it. The second asserts the nominal path stays quiet, which is what stops a reporting change from becoming noise that gets silenced again later.

ruff clean; targeted suite 74 passed / 2 skipped. All checks green, mergeable and clean.

Merging. With #136 already landed, the patch release follows.