Context (terra round, snapshot10g, gpt-6.1-terra/high): 8/8 pass, no product defects.
c1/c2: ARCHITECTURE.md covers 11/11 modules, all quoted identifiers exist; context-full NOT reached (limit 60000, reserve 9000, threshold 51000; peaks 46.5K/45.9K) because terra counts ~26% fewer tokens than Claude (scale 0.736). Mid-turn auto-compaction and the 'stop' choice were not exercised on terra.
c3: /compact 35,970 -> 9,354 tokens, driver counted 1; summary written in parts with working continuation notes; kept tail replayed its encrypted reasoning OK; no 400; turn 3 repeated no work. Stop path not exercised (no context-full).
c4: 3 context-full + 3 handoffs over 4 sessions at the right point (85% of 40000 incl. reserve 6000); original request quoted each time; interrupted messages saved with handoff reason; 63 requests summed across 4 sessions recompute to $0.30377345 exactly; no continuations carried across sessions; 56/56 continuations matched.
c5: REPORT.md counts verified (W20 69326 …); c6: INDEX.md 150/150 correct; c7: cache read 88.3% over 13 requests, prefix stable (system frozen, memory in user <context>); j6: 25/25 tests, cache read 94.4% over 30 requests, 30/30 continuations matched.
Known latent assistantItems issue: never triggered (127 continuations, none with 2+ text parts).
Notes: gpt-6.1-terra not in MODEL_CONTEXT_LIMITS -> 200K fallback (honestly labelled). Model wrote scratch under /tmp in bypass mode (j6). Per-part compaction summaries carry stale 'continue reading' lines (model).
