RF-18 — Snapshot-route telemetry prices the wrong artifact

State: fixed 2026-08-27 Repair shape: local telemetry correction
Severity: low

Finding

For (snapshot required) grounding, the worker is supposed to read the derived snapshot, but availability, size, and consumption telemetry charge the linked ingest path. The snapshot may be much larger, so the recorded context cost does not describe the artifact actually consumed.

Evidence

• resolve_note_markdown_links() intentionally prices the ingest while checking only snapshot availability.
• The prompt protocol tells workers to report the ingest path for this route.
• The former V1 regression in test_job_prompt.py fixed the mismatch as expected behavior before this repair.

Why it matters

Telemetry cannot support prompt-cost, reading-budget, or capacity decisions if the charged object is not the object read. The mismatch also obscures the exact evidence dependency described in RF-04.

Provisional repair direction

Resolve snapshot-required links to the logical ingest and both physical inputs the grounding check consumes: ingest metadata and derived snapshot bytes. Charge both files and require consumption metadata to name each one actually opened.

Done when

• Availability and consumption telemetry name the ingest and derived snapshot.
• Charged bytes equal both files' actual whole-file sizes.
• The logical ingest-to-physical-input routes remain explicit.
• The existing V1 mismatch test is replaced by the intended behavior.

Resolution

Resolved links now distinguish one link_target_path from one or more physical consumption targets. Ordinary routes map the path to itself. A (snapshot required) route maps the linked ingest to both the ingest and its derived snapshot, using each file's actual size. Availability schema v3 exposes both route mappings, separate logical and physical counts, and both priced artifacts. The generated prompt requires opened_paths to name each physical target the reviewer opened. Consumption schema v2 then prices those reported paths. Tests cover resolution, missing snapshots, prompt instructions, multi-target lineage telemetry, combined byte cost, and the logical-count/physical-cost boundary.