Picking up where epoch #79 left off. Triaged sprint:

• Implement Responses API: misleading error on context overflow, must communicate token limit exceeded #19 fix: Responses API truncation status — verified last cycle that to_json_oaicompat_resp{,_stream} still hardcode "status": "completed" regardless of stop == STOP_TYPE_LIMIT. The minimum fix per the issue: branch on stop type, emit "status": "incomplete" + "incomplete_details": {"reason": "max_output_tokens"} per OAI Responses spec. Test via the existing pytest harness with low --n-predict + check the JSON shape.
• Test coverage for PR feat(fit): expose per-device byte plan from common_fit_params (#66 step 1) #69 (fit per-device plan) — landed alongside PR feat(fit-params): --fit-print-plan emits per-device byte plan as JSON (#66 step 2 prep) #72/test(fit-params): smoke for --fit-print-plan JSON output #75 but the per-device plan output itself has no direct smoke test (PR test(fit-params): smoke for --fit-print-plan JSON output #75 covers the fit-params tool CLI surface, not the new out_bytes_per_device parameter directly). Add a focused C++ unit test or extend the fit-params smoke to assert plan-vs-status-success invariant.
• Bug-hunt sweep round 3 on tools/server/server-queue.cpp — round 1 found races in server-models.cpp (PR fix(server): guard two mapping[] accesses against silent default-insert #77), round 2 found UB in server-context.cpp create_checkpoint (folded into PR fix(server): per-slot byte cap on context checkpoints (closes #67) #68). Next adjacent file. Smaller than server-context.cpp; pattern continues to be lifetime + lock-scope.
• CI workflow audit round 2 — last epoch covered the zero-job-on-master pattern (build-cann.yml → PR ci(cann): drop auto-triggers (workflow has zero jobs upstream) #80). Other patterns worth scanning: self-hosted runner workflows that may not have runners in our fork, workflows with default tokens that need org-scoped permissions.
• Docs review: tools/server/README.md — PR docs(server): drop dead chat.{mjs,sh} links + fix embedding example path #76 fixed two broken links last epoch; the rest of the doc deserves a pass for stale examples, removed features, post-master-sync drift.

For later / saved:

• DFlash + parallel >1 implementation (Task chore(sync): upstream master → ht (114-commit sync 2026-06-13) #107) — gated on Router co-loads same-device-pinned models and OOMs: --models-max fit decision ignores per-device VRAM #66 step 5 + Markus signal
• Logit forward-parity test (Task Make a tag/release ggml-org/llama.cpp#111) — gated on getting PyTorch env
• Router co-loads same-device-pinned models and OOMs: --models-max fit decision ignores per-device VRAM #66 steps 3-5 (Task Won't compile on a MacBook Pro M1 8 GB ggml-org/llama.cpp#123) — awaiting Markus review of foundation PRs feat(fit): expose per-device byte plan from common_fit_params (#66 step 1) #69/feat(fit-params): --fit-print-plan emits per-device byte plan as JSON (#66 step 2 prep) #72
• Add Native Gemini API Compatibility to Server #5 (Gemini API compat) — bigger surface, future epoch
• feat: analytical fit for large MoE models (avoid repeated model loads) #11 (analytical fit for MoE) — depends on broader fit-params rework
• Steering hints: mid-inference context injection via overlapping activations #17 (steering hints) — research-level
• TBQ KV-cache (tbq3_0 / tbq4_0): 65-73x perplexity regression — recommend mark experimental #70 (TBQ KV-cache mark experimental) — PR chore(args): mark tbq3_0 / tbq4_0 KV-cache as experimental (closes #70) #71 is the action, awaiting Markus

Blocked on Markus / snoop-kube (unchanged):

• 10 open PRs awaiting review (fix(server): per-slot byte cap on context checkpoints (closes #67) #68, feat(fit): expose per-device byte plan from common_fit_params (#66 step 1) #69, chore(args): mark tbq3_0 / tbq4_0 KV-cache as experimental (closes #70) #71, feat(fit-params): --fit-print-plan emits per-device byte plan as JSON (#66 step 2 prep) #72, ci(sycl): drop auto-triggers (workflow has zero jobs upstream) #74, test(fit-params): smoke for --fit-print-plan JSON output #75, docs(server): drop dead chat.{mjs,sh} links + fix embedding example path #76, fix(server): guard two mapping[] accesses against silent default-insert #77, perf(dflash): gate top-5 logits diagnostic behind LLAMA_DFLASH_DEBUG #78, ci(cann): drop auto-triggers (workflow has zero jobs upstream) #80)
• Issue TBQ KV-cache (tbq3_0 / tbq4_0): 65-73x perplexity regression — recommend mark experimental #70 strategic decision
• Task fix(server): return 400 (not 500) for truncated tool-call args in input (#19) #109 titan rollout (b0daec5 image staged)
• Task chore: gitignore UI node_modules and .svelte-kit build artifacts #110 Q8 DFlash bench