Generalize Local Connector Bounded Reads

The Codex and Claude Code local collectors now avoid whole-file and whole-result-set memory spikes, but the same source-size-to-heap hazard remains in peer filesystem and local-DB connectors. Heavy local exports such as Twitter archives, iMessage chat.db, and Slack dump databases can be hundreds of megabytes or larger. A connector-wide contract is needed so bounded memory is not a one-off property of two agent connectors.

bound-codex-collector-memory fixed the incident class that caused a Codex collector to peak around 1.4-1.8 GB RSS. The same audit found the fix is complete for codex and claude_code, but the class-level problem remains:

The reference implementation SHALL collect filesystem-backed and local-database-backed connector state without retaining whole stream-eligible user files, whole unbounded local database result sets, or source-size-proportional parser state in process memory before record-level bounds apply. Connectors MAY read a whole file only when the file is a reviewed small per-artifact input or the code carries an explicit bounded exception with a reason. Local database connectors SHALL iterate unbounded user-data result sets row by row rather than materializing the full result in memory.