fix(eda): worker subscribed to a dead topic — async extractions never processed

[ancongui]

Problem

Queue-backed (POST /api/v1/extractions) extractions never completed. A submitted extraction sat in queued indefinitely; only the reaper revived it (~10 min later, after queued_orphan_threshold_s).

Root cause

The async worker's PostgresEventBus subscribes to pyfly.eda.destinations, whose default in pyfly.yaml was flydocs.jobs. But the submit handler publishes the extraction.submitted event to jobs_topic = flydocs.extractions (IDPSettings.jobs_topic, config.py). So the worker container subscribed to a topic nothing publishes to:

PostgresEventBus started: channel=flydocs_eda destinations=['flydocs.jobs'] group=flydocs-workers

while the outbox row was published to flydocs.extractions. The consumer offset never advanced.

Fix

Align the destinations default to flydocs.extractions (matching jobs_topic). Behaviour is unchanged for anyone who already sets FLYDOCS_EDA_DESTINATIONS, and the bbox-worker keeps its own flydocs.extractions.post_processing override + flydocs-bbox-workers group.

Also

• Bump the locked pyfly framework dependency to 26.6.99.
• Release 26.6.4 (version + CHANGELOG).

Verification

Reproduced and confirmed against a live stack (api + worker + bbox-worker on postgres EDA): before, a submitted extraction stayed queued with the worker idle; after, the worker immediately claims it (destinations=['flydocs.extractions']), runs the pipeline (load → classify → plan_tasks → extract → …) and delivers the result webhook.