fix(server): restore graph_event_stream module forwarders, event_meta included

[cdeust]

Summary

Fixes #134. Commit 45d4a80 deleted graph_event_stream's module-level
emit/close/reset forwarders on the premise they had never had a caller
in this repository's history. That premise was false: graph_build_run.py
(reset(), emit(), the finally-block close()) and
graph_build_merge.py (emit()) all call them — the module is imported
once and the forwarders used for the lifetime of a build. The build now
dies on its first statement with AttributeError, and /api/graph/progress
never advances past starting.

The commit's other stated reason is the real defect: emit() grew an
event_meta parameter and the forwarder silently dropped it. That is a bug
in the forwarder — it was used as the reason to delete it instead, taking
every caller down with it. The module is also published API, so removing
public functions in a patch-level move breaks external consumers too.

What changed

• Restored all three forwarders, functional: emit forwards every
  parameter including event_meta, close/reset delegate to the process
  singleton via get_stream(), both build-side callers unchanged.
• Removed the "deliberately no module-level forwarders" comment — it
  documented a decision that is now withdrawn.

Tests (tests/test_graph_event_stream_forwarders.py)

• Each forwarder delegates to the process singleton returned by
  get_stream().
• emit's forwarder carries event_meta through to what a subscriber
  receives — the exact drift that started the incident. Verified failing
  against the pre-45d4a80 forwarder and passing against this one.
• The real production binding: imports the module once and drives
  reset/emit/close as module-level calls, the way
  graph_build_run.py actually does — never touching get_stream()
  directly, the way every existing merge test's fake stream does. This is
  the binding that broke and no prior test exercised it.

Verification

• uv run pytest tests/ -k "graph_event or graph_build or activity_stream or http_live" — 77 passed, 2 skipped.
• uv run ruff check / ruff format --check — clean.
• End-to-end: standalone server against a live Cortex DB. /api/graph/progress
  advances starting → loading … → baseline_ready → L6 sweep → layout bake → full_ready (224240 nodes / 652486 edges), zero
  background build error in the log.

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

[cdeust]

ZETETIC-REVIEW: APPROVE

This repairs a crash that shipped in 3.1.0, and the manner of the original break is the lesson. Commit 45d4a80 deleted emit/close/reset from graph_event_stream, justified in the code as: they "had never had a caller in this repository's history". They had four. The graph build died at its first statement, /api/graph/progress sat at starting forever, and Graph, Brain, Knowledge, Board and Wiki stayed empty.

The commit's other stated motive — that emit had grown an event_meta parameter the forwarder silently dropped — was the actual defect and the actual reason to repair rather than delete. A drifted forwarder is a bug in the forwarder. Using it as grounds for removal left the drift unfixed and took four working callers with it.

The premise was checked against the primary artifact before anything was touched, which is the discipline whose absence caused this: the module was read on origin/main and its callers grepped and then read — graph_build_run.py:155, :234, :428-433, :547-549, and graph_build_merge.py:136. That is one more call site than the issue itself listed.

The forwarders are restored plugged, not routed around. emit forwards event_meta; close and reset delegate to the singleton; both callers are untouched; the "deliberately no module-level forwarders" comment is gone, since it documented a decision that has been withdrawn. Restoring three public functions also undoes an unannounced removal of published API in a patch-level move.

The test that matters is the one that would have caught this, and it is here. Every existing merge test injects a SimpleNamespace(emit=MagicMock()), which is exactly why none of them saw the defect: the fault was in how the caller obtains the stream, not in what it does with it. test_production_build_binding_reaches_the_real_singleton drives the module the way graph_build_run.py does, so a call site and the module can no longer drift apart unobserved.

Proven to fail against the pre-fix code, not merely to pass against the new. Stashed back to the pre-45d4a80 forwarder shape: all five new tests fail. Restored: 12 pass, the five new plus the seven pre-existing meta tests.

The acceptance signal is external and end-to-end. The standalone server was launched against a live Cortex Postgres and /api/graph/progress advanced starting → loading → baseline_ready → L6 sweep → layout (DrL bake) → full_ready, finishing at 224,240 nodes and 652,486 edges with zero background build error. A green unit suite would not have been sufficient here — the whole defect was that unit tests passed while the product did nothing.

ruff check and format --check clean; targeted suite 77 passed / 2 skipped, no regressions. All checks green, mergeable and clean.

Merging. A patch release follows once #137 lands, and I will send the landing sha to the session that pins this plugin.