Control channel: scope buffer page reads to referenced pool entries and attachments

[ransomr]

This PR contains:

• New features
• Changes to dev-tools e.g. CI config / github tooling
• Docs
• Bug fixes
• Code refactor

What is the current behavior? (You can also link to an open issue here)

Fixes meridianlabs-ai#80

When a control-channel events page is served from the realtime sample buffer (open_sample_history_from → _sample_history), the page limit only bounds the event rows. Each page request also unconditionally loads and json.loads the sample's entire message pool and call pool, and fetches every attachment body for the sample (tens of MB on image-heavy transcripts). Since the read runs synchronously on the eval's own event loop, a read-only monitor polling inspect ctl events imposes repeated blocking pauses on in-flight samples.

What is the new behavior?

Fix option 1 from the issue — page-scoped pool/attachment loading:

• SampleHistory pools are now position-keyed mappings (dict[int, ...]) rather than dense lists. Full-history reads carry every entry (contiguous keys from 0); page reads carry only referenced positions.
• After fetching a page's event rows, _sample_history(page_scoped=True) collects the input_refs/call_refs positional ranges and the attachment:// hashes referenced by the page's event JSON (plus refs inside the loaded pool entries, since large message content is itself stored as an attachment), and loads only those rows (WHERE pos IN (...) via ROW_NUMBER rank, WHERE hash IN (...) scoped to the sample; IN lists chunked under SQLite's bound-parameter limit).
• Both paged entry points (open_sample_history_from and open_sample_history_tail) are page-scoped; open_sample_history (full reads: streaming flush, recorder, retry errors) still loads complete pools.
• _expand_refs / resolve_model_event_inputs / resolve_model_event_calls / materialize_pooled_events accept position-keyed mappings alongside dense lists; the buffer transcript history provider resolves against the mappings.
• SampleHistory.events_data (dense pools for .eval serialization) became a property that raises on a page-scoped (sparse) history rather than silently misaligning refs.

Does this PR introduce a breaking change? (What changes might users need to make in their application due to this PR?)

No — all changed interfaces are internal (inspect_ai.log._recorders.buffer.*, inspect_ai.event._pool). Public resolve_* helpers only widened their accepted parameter types.

Other information:

Verification: pytest tests/log tests/checkpoint tests/_control tests/_eval/test_retry_error_events.py (all passing), ruff check/ruff format, and mypy clean on changed files. New tests in tests/log/test_sample_history.py cover page-scoped message/call pool loading, attachment scoping (event-level and pool-entry-level), end-to-end materialization through BufferTranscriptHistoryProvider.events_from, and the events_data guard.

The keyset-cursor option (#2 in the issue) is not included; the ROW_NUMBER CTE still scans the sample's event rows per page, but no longer parses or ships unreferenced pool/attachment content.

🤖 Generated with Claude Code