feat: compaction boundary-range columns + effective-context derivation

[Sinity]

Add session_events.boundary_start_position/boundary_end_position (message range a compaction replaces, in the session's own position coordinate) + boundary_message_id (the materialized summary). Parser computes the range as it walks records (start = prev boundary end + 1; end = message_position-1); writer applies position_offset. Add read helper get_effective_context(session, at_position): returns [summary] + post-boundary (what the model actually saw) vs the full composed prefix used for forks. Schema bump (v14->v15), re-ingest. Surfaces: MCP/CLI 'view=effective_context', pathology stale_context gets a precise replaced-range signal. Full design in agent report. Ref #2467.

[Sinity]

Folded into #2467 for consolidated lineage tracking (compaction boundary-range schema slice). Kept open and labeled for the actionable feature.

[Sinity]

2026-07-26 current-reality reconciliation

Audited against master at b14ec34e and current lineage code.

Disposition: keep open. This is a real residual, not stale umbrella debt.

The main lineage normalization and iterative transcript composition work has landed, including fork/resume/compaction prefix handling and the subagent-versus-main agent-acompact-* distinction. However, production code still has no boundary_start_position / boundary_end_position fields or equivalent precise replaced-range model; those names appear only in planning/handoff material.

Current consequence: Polylogue can compose logical history and classify compaction relationships, but cannot yet answer the narrower question “what effective context did the model actually see after this compaction boundary?” with a precise replaced range.

Close when:

• the parser records evidence-backed compaction boundary ranges;
• writer position offsets and lineage composition preserve those coordinates;
• CLI/MCP/API expose an effective-context read distinct from full logical transcript;
• tests cover repeated compaction, subagent self-compaction, and stale/ambiguous boundaries;
• a real archive query demonstrates the result.

Do not infer boundaries from summary prose alone.

[Sinity]

Tracker reconciliation: this GitHub issue is a mirror of Bead polylogue-4ts.5 (tracker:gh-mirror). Both own precise compaction boundary ranges and effective-context derivation; their scope and state should remain synchronized.